Contoh Makalah mengenai Sinapsis. Total input currently arising out of incoming signals at synapses

 

 

Deep Sparse Rectifier Neural Networks

Xavier Glorot, Antoine Bordes and Yoshua Bengio

Dept. IRO

University of Montreal

Montreal, QC

glorotxa,bordesa,bengioy@iro.umontreal.ca

Abstract

Rectifying neurons are more biologically plausible than logistic sigmoid neurons, which are themselves more biologically plausible than hyperbolic tangent neu¬rons. However, the latter work better for training multi-layer neural networks than logistic sigmoid neurons. This paper shows that networks of rectifying neu¬rons yield equal or better performance than hyperbolic tangent networks in spite of the hard non-linearity and non-differentiability at zero and create sparse rep¬resentations with true zeros which are remarkably suitable for naturally sparse data. Even though they can take advantage of semi-supervised setups with extra-unlabeled data, deep rectifier networks can reach their best performance without requiring any unsupervised pre-training on purely supervised tasks with large la¬beled datasets. Hence, these results can be seen as a new milestone in the attempts at understanding the difficulty in training deep but purely supervised neural net¬works, and closing the performance gap between neural networks learnt with and without unsupervised pre-training.

1 Introduction

Many differences exist between the neural network models used by machine learning researchers and those used by computational neuroscientists. This is in part because the objective of the former is to obtain computationally efficient learners, that generalize well to new examples, whereas the objective of the latter is to abstract out neuroscientific data while obtaining explanations of the prin¬ciples involved, providing predictions and guidance for future biological experiments. Areas where both objectives coincide are therefore particularly worthy of investigation, pointing towards compu¬tationally motivated principles of operation in the brain that can also enhance research in artificial intelligence. In this paper we show that two common gaps between computational neuroscience models and machine learning neural network models can be bridged by using the following linear by part activation: max(0, x), called the rectifier (or hinge) activation function. Experimental results will show engaging training behavior of this activation function, especially for deep architectures (see Bengio (2009) for a review), i.e., where the number of hidden layers is 3 or more.

Recent theoretical and empirical work in statistical machine learning has demonstrated the impor¬tance of learning algorithms for deep architectures. This is in part inspired by observations of the mammalian visual cortex, which consists of a chain of processing elements, each of which is asso¬ciated with a different representation of the raw visual input. This is particularly clear in the primate visual system (Serre et al., 2007), with its sequence of processing stages: detection of edges, primi¬tive shapes, and moving up to gradually more complex visual shapes. Interestingly, it was found that the features learned in deep architectures resemble those observed in the first two of these stages (in areas V1 and V2 of visual cortex) (Lee et al., 2008), and that they become increasingly invariant to factors of variation (such as camera movement) in higher layers (Goodfellow et al., 2009).

 

1

 

 

Figure 1: Left: common activation function motivated by biological data. Right: commonly used activation functions in neural networks literature: logistic sigmoid and hyperbolic tangent (tanh).

Regarding the training of deep networks, something that can be considered a breakthrough happened in 2006, with the introduction of Deep Belief Networks (Hinton et al., 2006), and more generally the idea of initializing each layer by unsupervised learning (Bengio et al., 2007; Ranzato et al., 2007). Some authors have tried to understand why this unsupervised procedure helps (Erhan et al., 2010) while others investigated why the original training procedure for deep neural networks failed (Ben-gio and Glorot, 2010). From the machine learning point of view, this paper brings additional results in these lines of investigation.

We propose to explore the use of rectifying non-linearities as alternatives to the hyperbolic tangent or sigmoids in deep artificial neural networks, in addition to using an L1 regularizer to promote spar¬sity and prevent potential numerical problems with unbounded activation. Nair and Hinton (2010) present promising results of the influence of such units in the context of Restricted Boltzmann Ma¬chines compared to logistic sigmoid activations. Our work extends this for the case of pre-training using denoising auto-encoders (Vincent et al., 2008) and provides an extensive empirical comparison of the rectifying activation function against the hyperbolic tangent on image classification bench¬marks as well as an original derivation for the text application of sentiment analysis.

Our experiments indicate that training proceeds better when the artificial neurons are either off or operating mostly in a linear regime. Surprisingly, rectifying activation allows deep networks to achieve their best performance even without unsupervised pre-training. Hence, our work proposes a new contribution to the trend of understanding and merging the performance gap between deep networks learnt with and without unsupervised pre-training (Erhan et al., 2010; Bengio and Glorot, 2010). Interestingly, rectifier networks can still benefit from unsupervised pre-training in the context of semi-supervised learning where large amounts of unlabeled data are provided. Furthermore, as rectifier units naturally lead to sparse networks and are closer to biological neurons’ responses in their main operating regime, this work also bridges (in part) a machine learning / neuroscience gap in terms of activation function and sparsity.

2 Background

2.1 Neuroscience Observations

For models of biological neurons, the activation function is the expected firing rate as a function of the total input currently arising out of incoming signals at synapses (Dayan and Abott, 2001). An activation function is termed, respectively antisymmetric or symmetric when its response to the opposite of a strongly excitatory input pattern is respectively a strongly inhibitory or excitatory one, and one-sided when this response is zero. The main gaps that we wish to consider between computational neuroscience models and machine learning models are the following.

Studies on brain energy expense suggest that neurons encode information in a sparse and distributed way (Attwell and Laughlin, 2001), estimating the percentage of neurons active at the same time to be between 1 and 4% (Lennie, 2003). This corresponds to a trade-off between richness of represen¬tation and small action potential energy expenditure. Without additional regularization, such as an L1 penalty, ordinary feedforward neural nets do not have this property. For example, the sigmoid activation has a steady state regime around 21, therefore, after initializing with small weights, all neu 

2

 

rons fire at half their saturation regime. This is biologically implausible and hurts gradient-based optimization (LeCun et al., 1998; Bengio and Glorot, 2010).

Important divergences between biological and machine learning models concern non-linear activa¬tion functions. The leaky integrate-and-fire (or LIF) (Dayan and Abott, 2001), gives the following relation between the firing rate and the input current, illustrated in Figure 1 (left):

 

 

where tref is the refractory period (minimal time between two action potentials), I the input current, Vr the resting potential and Vth the threshold potential (with Vth > Vr), and R, E, τ the membrane resistance, potential and time constant. The most commonly used activation functions in the deep learning and neural networks literature are the standard logistic sigmoid and the hyperbolic tangent (see Figure 1, right). The hyperbolic tangent has a steady state at 0, and is therefore preferred from the optimization standpoint (LeCun et al., 1998; Bengio and Glorot, 2010), but it forces an antisymmetry around 0 which is absent in biological neurons.

2.2 Advantages of Sparsity

Sparsity has become a concept of interest, not only in computational neuroscience and machine learning but also in statistics and signal processing (Candes and Tao, 2005). It was first introduced in computational neuroscience in the context of sparse coding in the visual system (Olshausen and Field, 1997). It has been a key element of deep convolutional networks exploiting a variant of auto-encoders (Ranzato et al., 2007, 2008; Mairal et al., 2009) with a sparse distributed representation, and has also become a key ingredient in Deep Belief Networks (Lee et al., 2008). A sparsity penalty has been used in several computational neuroscience (Olshausen and Field, 1997; Doi et al., 2006) and machine learning models (Lee et al., 2007; Mairal et al., 2009), in particular for deep architec¬tures (Lee et al., 2008; Ranzato et al., 2007, 2008). However, in the latter, the neurons end up taking small but non-zero activation or firing probability. We show here that using a rectifying non-linearity gives rise to real zeros of activations and thus truly sparse representations. From a computational point of view, such representations are appealing for the following reasons:

• Information disentangling. One of the claimed objectives of deep learning algorithms (Bengio, 2009) is to disentangle the factors explaining the variations in the data. A dense representation is highly entangled because almost any change in the input modifies most of the entries in the representation vector. Instead, if a representation is both sparse and robust to small input changes, the set of non-zero features is almost always roughly conserved by small changes of the input.

• Efficient variable-size representation. Different inputs may contain different amounts of infor¬mation and would be more conveniently represented using a variable-size data-structure, which is common in computer representations of information. Varying the number of active neurons allows a model to control the effective dimensionality of the representation for a given input and the required precision.

• Linear separability. Sparse representations are also more likely to be linearly separable, or more easily separable with less non-linear machinery, simply because the information is represented in a high-dimensional space. Besides, this can reflect the original data format. In text-related applications for instance, the original raw data is already very sparse (see Section 4.2).

• Distributed but sparse. Dense distributed representations are the richest representations, being potentially exponentially more efficient than purely local ones (Bengio, 2009). Sparse represen¬tations’ efficiency is still exponentially greater, with the power of the exponent being the number of non-zero features. They may represent a good trade-off with respect to the above criteria.

Nevertheless, forcing too much sparsity may hurt predictive performance for an equal number of neurons, because it reduces the effective capacity of the model.

3

 

 

Figure 2: Left: sparse propagation of activations and gradients in a network of rectifier units. The input selects a subset of active neurons and computation is linear in this subset. Right: The rectifier and the softplus activation functions, the second one is a smooth version of the first.

3 Deep Rectifier Networks

3.1 Rectifier Neurons

The neuroscience literature (Bush and Sejnowski, 1995; Douglas and al., 2003) indicates that corti¬cal neurons are rarely in their maximum saturation regime, and suggests that their activation func¬tion can be approximated by a rectifier. Most previous studies of neural networks involving a recti¬fying activation function concern recurrent networks (Salinas and Abbott, 1996; Hahnloser, 1998). The rectifier function rectifier(x) = max(0, x) is one-sided and therefore does not enforce a sign symmetry1 or antisymmetry1: instead, the response to the opposite of an excitatory input pattern is 0 (no response). However, we can obtain symmetry or antisymmetry by combining two rectifier units sharing parameters.

Advantages The rectifier activation function allows a network to easily obtain sparse represen¬tations. For example, after uniform initialization of the weights, around 50% of hidden units con¬tinuous output values are real zeros, and this fraction can easily increase with sparsity-inducing regularization. Apart from being more biologically plausible, sparsity also leads to mathematical advantages (see previous section).

As illustrated in Figure 2 (left), the only non-linearity in the network comes from the path selection associated with individual neurons being active or not. For a given input only a subset of neurons are active. Computation is linear on this subset: once this subset of neurons is selected, the output is a linear function of the input (although a large enough change can trigger a discrete change of the active set of neurons). The function computed by each neuron or by the network output in terms of the network input is thus linear by parts. We can see the model as an exponential number of linear models that share parameters Nair and Hinton (2010). Because of this linearity, gradients flow well on the active paths of neurons (there is no gradient vanishing effect due to activation non-linearities of deep networks of sigmoid or tanh units), and mathematical investigation is easier. Computations are also cheaper: there is no need for computing the exponential function in activations, and sparsity can be exploited.

Potential Problems One may hypothesize that the hard saturation at 0 may hurt optimization by blocking gradient back-propagation. To evaluate the potential impact of this effect we also investi¬gate the softplus activation: softplus(x) = log(1 + ex) (Dugas et al., 2001), a smooth version of the rectifying non-linearity. We lose the exact sparsity, but may hope to gain easier training. However, experimental results (see Section 4.1) tend to contradict that hypothesis, suggesting that hard zeros can actually help supervised training. We hypothesize that the hard non-linearities do not hurt so long as the gradient can propagate along some paths, i.e., that some of the hidden units in each layer are non-zero. With the credit and blame assigned to these ON units rather than distributed

1The hyperbolic tangent absolute value non-linearity  tanh(x) used by Jarrett et al. (2009) enforces sign symmetry. A tanh(x) non-linearity enforces sign antisymmetry.

 

4

 

more evenly, we hypothesize that optimization is easier. Another problem could arise due to the unbounded behavior of the activations; one may thus want to use a regularizer to prevent potential numerical problems. Therefore, we use the L1 penalty on the activation values, which also pro¬motes additional sparsity. Also recall that, in order to efficiently represent symmetric/antisymmetric behavior in the data, a rectifier network would need twice as many hidden units as a network of symmetric/antisymmetric activation functions.

Finally, rectifier networks are subject to ill-conditioning of the parametrization. Biases and weights can be scaled in different (and consistent) ways while preserving the same overall network function. More precisely, consider for each layer of depth i of the network a scalar αi, and scaling the parame-s

ters as W~i = Wi

αi and b~i = bi 

H

i . The output units values then change as follow: s' = ~n

j=1 αj j=1 αj .

Therefore, as long as jnj=1 αj is 1, the network function is identical.

3.2 Unsupervised Pre-training

This paper is particularly inspired by the sparse representations learned in the context of auto-encoder variants, as they have been found to be very useful in training deep architectures (Bengio, 2009), especially for unsupervised pre-training of neural networks (Erhan et al., 2010).

Nonetheless, certain difficulties arise when one wants to introduce rectifier activations into stacked denoising auto-encoders (Vincent et al., 2008). First, the hard saturation below the threshold of the rectifier function is not suited for the reconstruction units. Indeed, whenever the network happens to reconstruct a zero in place of a non-zero target, the reconstruction unit can not backpropagate any gradient.2 Second, the unbounded behavior of the rectifier activation also needs to be taken into account. We have experimented with several strategies, and propose the following two (the first strategy has proven to be the most efficient on image experiments and the second one on text ones):

1. Use a softplus activation function for the reconstruction layer, along with a quadratic re¬construction cost.

2. Scale the rectifier activation values coming from the previous encoding layer to bound them between 0 and 1, then use a sigmoid activation function for the reconstruction layer, along with a cross-entropy reconstruction cost.

4 Experimental Study

4.1 Image Recognition

Experimental setup We considered the image datasets detailed below. Each of them has a train¬ing set (for tuning parameters), a validation set (for tuning hyper-parameters) and a test set (for reporting generalization performance). They are presented according to their number of train-ing/validation/test examples, their respective image sizes, as well as their number of classes:

• MNIST (LeCun et al., 1998): 50k/10k/10k, 28 × 28 digit images, 10 classes.

• CIFAR10 (Krizhevsky and Hinton, 2009): 50k/5k/5k, 32 × 32 × 3 RGB images, 10 classes.

• NISTP: 81920k/80k/20k, 32 × 32 character images from the NIST database 19, with randomized distortions (Anonymous, 2010), 62 classes. This dataset is much larger and more difficult than the original NIST (Grother, 1995).

• NORB: 233172/58428/58320, taken from Jittered-Cluttered NORB (LeCun et al., 2004). Stereo-pair images of toys on a cluttered background, 6 classes. The data has been preprocessed similarly to (Nair and Hinton, 2010): we subsampled the original 2 × 108 × 108 stereo-pair images to 2 × 32 × 32 and scaled linearly the image in the range [1,1]. We followed the procedure used by Nair and Hinton (2010) to create the validation set.

For all experiments except on the NORB data (LeCun et al., 2004), the models we used are stacked denoising auto-encoders (Vincent et al., 2008) with three hidden layers and 1000 units per layer.

2Why is this not a problem for hidden layers too? we hypothesize that it is because gradients can still flow through the active (non-zero), possibly helping rather than hurting the assignment of credit.

 

5

 

Table 1: Test error on networks of depth 3. Bold results represent statistical equivalence between similar experiments, with and without pre-training, under the null hypothesis of the pairwise test with P = 0.05.

Neuron MNIST CIFAR10 NISTP NORB

With unsupervised pre-training

Rectifier 1.20% 49.96% 32.86% 16.46%

Tanh 1.16% 50.79% 35.89% 17.66%

Softplus 1.17% 49.52% 33.27% 19.19%

Without unsupervised pre-training

Rectifier 1.43% 50.86% 32.64% 16.40%

Tanh 1.57% 52.62% 36.46% 19.29%

Softplus 1.77% 53.20% 35.48% 17.68%

The architecture of Nair and Hinton (2010) has been used on NORB: two hidden layers with respec¬tively 4000 and 2000 units. We used a cross-entropy reconstruction cost for tanh networks and a quadratic cost over a softplus reconstruction layer for the rectifier and softplus networks. We chose masking noise as the corruption process: each pixel has a probability of 0.25 of being artificially set to 0. The unsupervised learning rate is constant, and the following values have been explored: .1, .01, .001, .0001. We select the model with the lowest reconstruction error. For the super¬vised fine-tuning we chose a constant learning rate in the same range as the unsupervised learning rate with respect to the supervised validation error. The training cost is the negative log likelihood  log P (correct classinput) where the probabilities are obtained from the output layer (which im¬plements a softmax logistic regression). We used stochastic gradient descent with mini-batches of size 10 for both unsupervised and supervised training phases.

To take into account the potential problem of rectifier units not being symmetric around 0, we use a variant of the activation function for which half of the units output values are multiplied by -1. This serves to cancel out the mean activation value for each layer and can be interpreted either as inhibitory neurons or simply as a way to equalize activations numerically. Additionally, an L1 penalty on the activations with a coefficient of 0.001 was used during pre-training and fine-tuning in order to increase the amount of sparsity in the learned representations.

Main results Table 1 summarizes the results on networks of 3 hidden layers of 1000 hidden units each, comparing all the neuron types on all the datasets, with or without unsupervised pre-training. In the latter case, the supervised training phase has been carried out using the same experimental setup as the one described above for fine-tuning. The main observations we make are the following:

• Despite the hard threshold at 0, networks trained with the rectifier activation function can find local minima of greater or equal quality than those obtained with its smooth counterpart, the softplus. Rectifiers are not only biologically plausible, they are also computationally efficient.

• There is almost no improvement when using unsupervised pre-training with rectifier activations, contrary to what is experienced using tanh or softplus. Purely supervised rectifier networks re¬main competitive on all 4 datasets, even against the pretrained tanh or softplus models.

• Rectifier networks are truly deep sparse networks. There is an average exact sparsity (fraction of zeros) of the hidden layers of 83.40% on MNIST and 72.00% on CIFAR10. Figure 3 (left) provides a better understanding of the influence of sparsity. It displays the MNIST test error of deep rectifier networks (without pre-training) according to different average sparsity obtained by varying the L1 penalty on the activations. Networks appear to be quite robust to it as models with 70% to almost 85% of true zeros can achieve similar performances.

With labeled data, deep rectifier networks appear to be attractive models. They are biologically cred¬ible, and, compared to their standard counterparts, do not seem to depend as much on unsupervised pre-training, while ultimately yielding sparse representations.

This last conclusion is slightly different from those reported in (Nair and Hinton, 2010) in which is demonstrated that unsupervised pre-training with Restricted Boltzmann Machines and using rectifier units is beneficial. In particular, the paper reports that pre-trained rectified Deep Belief Networks can achieve a test error on NORB below 16%. However, we believe that our results are compatible with those: we extend the experimental framework to a different kind of models (stacked denoising auto-encoders) and different datasets (on which conclusions seem to be different). Furthermore,

 

6

 

Figure 3: Left: influence of final sparsity on accuracy. 200 randomly initialized deep rectifier networks were trained on MNIST with various L1 penalties (from 0 to 0.001) to obtain different sparsity levels. Right: effect of unsupervised pre-training. On NORB, we compare hyperbolic tangent and rectifier networks, with or without unsupervised pre-training, and fine-tune only on subsets of increasing size of the training set.

note that our rectified model without pre-training on NORB is very competitive (16.4% error) and outperforms the 17.6% error of the non-pretrained model from Nair and Hinton (2010), which is basically what we find with the non-pretrained softplus units (17.68% error).

Semi-supervised setting Figure 3 (right) presents results of semi-supervised experiments con¬ducted on the NORB dataset. We vary the percentage of the original labeled training set which is used for the supervised training phase of the rectifier and hyperbolic tangent networks and evaluate the effect of the unsupervised pre-training (using the whole training set, unlabeled). Confirming conclusions of Erhan et al. (2010), the network with hyperbolic tangent activations improves with unsupervised pre-training for any labeled set size (even when all the training set is labeled). How¬ever, the picture changes with rectifying activations. In semi-supervised setups (with few labeled data), the pre-training is highly beneficial. But the more the labeled set grows, the closer the models with and without pre-training. Eventually, when all available data is labeled, the two models achieve identical performance. Rectifier networks can maximally exploit labeled and unlabeled information.

4.2 Sentiment Analysis

In text-related applications, data is usually very sparse. Deep rectifier networks, thanks to their naturally sparse behavior, could be an interesting match for this kind of learning task. To validate this intuition, we consider sentiment analysis, a text mining area which aims to determine the judgment of a writer with respect to a given topic (see (Pang and Lee, 2008) for a review). The basic task consists in classifying the polarity of reviews either by predicting whether the expressed opinions are positive or negative, or by assigning them star ratings on either 3, 4 or 5 star scales.

Following a task originally proposed by Snyder and Barzilay (2007), our data consists of restaurant reviews which have been extracted from the restaurant review site www.opentable.com. We have access to 10,000 labeled and 300,000 unlabeled training reviews, while the test set contains 10,000 examples. The goal is to predict the rating on a 5 star scale and performance is evaluated using Root Mean Squared Error (RMSE).3 The review text is treated as a bag of words and trans¬formed into binary vectors encoding the presence/absence of terms. For computational reasons, only the 5000 most frequent terms of the vocabulary are kept in the feature set.4 The resulting pre¬processed data is very sparse: 0.6% of non-zero features on average. Unsupervised pre-training of the networks employs both labeled and unlabeled training reviews while the supervised fine-tuning phase is carried out by 10-fold cross-validation on the labeled training examples.

The model are stacked denoising auto-encoders, with 1 or 3 hidden layers of 5000 hidden units and rectifier or tanh activation, which are trained in a greedy layer-wise fashion. Predicted ratings are defined by the expected star value computed using multiclass (multinomial, softmax) logistic

3Even though our tasks are identical, our database is larger than the one of (Snyder and Barzilay, 2007). 4Preliminary experiments suggested that larger vocabulary sizes did not markedly change results.

7

 

regression output probabilities. For rectifier networks, when a new layer is stacked, activation values of the previous layer are scaled within the interval [0,1] and a sigmoid reconstruction layer with a cross-entropy cost is used. We also add an L1 penalty during pre-training and fine-tuning. Because of the binary input, we use a “salt and pepper noise” (i.e. masking some inputs by zeros and others by ones) for unsupervised training of the first layer. A simple zero masking (as in (Vincent et al., 2008)) is used for the higher layers. We selected the noise level of each layer based on the classification performance, the other hyperparameters are selected according to the reconstruction error.

Table 2: Test RMSE and sparsity level obtained by 10-fold cross-validation on OpenTable data.

Network No hidden layer Rectifier (1-layer) Rectifier (3-layers) Tanh (3-layers)

RMSE 0.885 f 0.006 0.807 f 0.004 0.746 f 0.004 0.774 f 0.008

Sparsity 99.4% 28.9% f 0.2 53.9% f 0.7 00.0% f 0.0

Results are displayed in Table 2. Interestingly, the RMSE significantly decreases as we add hidden layers to the rectifier neural net. These experiments confirm that rectifier networks improve after an unsupervised pre-training phase in a semi-supervised setting: with no pre-training, the 3-layers model can not obtain a RMSE lower than 0.833. Additionally, although we can not replicate the original very high degree of sparsity of the training data, the 3-layers network can still attain an overall sparsity of more than 50%. Finally, on data with these particular properties (binary, high sparsity), the 3-layers network with tanh activation function (which has been learnt with the ex¬act same setup) is clearly outperformed. The sparse behavior of the deep rectifier network seems particularly suitable in this case, because the raw input is very sparse and varies in its number of non-zeros. The latter can also be achieved with sparse internal representations, not with dense ones.

5 Conclusion

Sparsity and neurons operating mostly in a linear regime can be brought together in more biologi¬cally plausible deep neural networks. Rectifier units help to bridge the gap between unsupervised pre-training and no pre-training, which suggests that they may help in finding better minima during training. This finding has been verified for four image classification datasets of different scales and all this in spite of their inherent problems, such as zeros in the gradient, or ill-conditioning of the parametrization. Rather sparse networks are obtained (from 50 to 80% sparsity for the best general¬izing models, whereas the brain is hypothesized to have 95% to 99% sparsity), which may explain some of the benefit of using rectifiers. Furthermore, rectifier activation functions have shown to be remarkably adapted to sentiment analysis, a text-based task with a very large degree of data sparsity. This promising result tends to indicate that deep sparse rectifier networks might yield powerful text mining tools in the future.

Acknowledgments

The authors are greatful for the support received from the following organizations: DARPA, NSERC, CBIE, CIFAR, the Canada Research Chairs, and MITACS.

References

Anonymous, A. (2010). Anonymous title. Technical report, Anonymous University.

Attwell, D. and Laughlin, S. (2001). An energy budget for signaling in the grey matter of the brain. Journal of Cerebral Blood Flow and Metabolism, 21(10), 1133–1145.

Bengio, Y. (2009). Learning deep architectures for AI. Foundations and Trends in Machine Learning, 2(1), 1–127. Also published as a book. Now Publishers, 2009.

Bengio, Y. and Glorot, X. (2010). Understanding the difficulty of training deep feedforward neural networks. In Proceedings of AISTATS 2010, volume 9, pages 249–256.

Bengio, Y., Lamblin, P., Popovici, D., and Larochelle, H. (2007). Greedy layer-wise training of deep networks. In NIPS 19, pages 153–160. MIT Press.

Bush, P. C. and Sejnowski, T. J. (1995). The cortical neuron. Oxford university press.

Candes, E. and Tao, T. (2005). Decoding by linear programming. IEEE Transactions on Information Theory, 51(12), 4203–4215.

 

8

 

Dayan, P. and Abott, L. (2001). Theoretical neuroscience. MIT press.

Doi, E., Balcan, D. C., and Lewicki, M. S. (2006). A theoretical analysis of robust coding over noisy overcom-plete channels. In NIPS’05, pages 307–314. MIT Press, Cambridge, MA.

Douglas, R. and al. (2003). Recurrent excitation in neocortical circuits. Science, 269(5226), 981–985.

Dugas, C., Bengio, Y., Belisle, F., Nadeau, C., and Garcia, R. (2001). Incorporating second-order functional knowledge for better option pricing. In NIPS 13. MIT Press.

Erhan, D., Bengio, Y., Courville, A., Manzagol, P.-A., Vincent, P., and Bengio, S. (2010). Why does unsuper¬vised pre-training help deep learning? JMLR, 11, 625–660.

Goodfellow, I., Le, Q., Saxe, A., and Ng, A. (2009). Measuring invariances in deep networks. In NIPS’09, pages 646–654.

Grother, P. (1995). Handprinted forms and character database, NIST special database 19. In National Institute of Standards and Technology (NIST) Intelligent Systems Division (NISTIR).

Hahnloser, R. L. T. (1998). On the piecewise analysis of networks of linear threshold neurons. Neural Netw., 11(4), 691–697.

Hinton, G. E., Osindero, S., and Teh, Y. (2006). A fast learning algorithm for deep belief nets. Neural Computation, 18, 1527–1554.

Jarrett, K., Kavukcuoglu, K., Ranzato, M., and LeCun, Y. (2009). What is the best multi-stage architecture for object recognition? In Proc. International Conference on Computer Vision (ICCV’09). IEEE.

Krizhevsky, A. and Hinton, G. (2009). Learning multiple layers of features from tiny images. Technical report, University of Toronto.

LeCun, Y., Bottou, L., Orr, G., and Muller, K. (1998). Efficient backprop.

LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. (1998). Gradient-based learning applied to document recog¬nition. Proceedings of the IEEE, 86(11), 2278–2324.

LeCun, Y., Huang, F.-J., and Bottou, L. (2004). Learning methods for generic object recognition with invariance to pose and lighting. In Proc. CVPR’04, volume 2, pages 97–104, Los Alamitos, CA, USA. IEEE Computer Society.

Lee, H., Battle, A., Raina, R., and Ng, A. (2007). Efficient sparse coding algorithms. In NIPS’06, pages 801–808. MIT Press.

Lee, H., Ekanadham, C., and Ng, A. (2008). Sparse deep belief net model for visual area V2. In NIPS’07, pages 873–880. MIT Press, Cambridge, MA.

Lennie, P. (2003). The cost of cortical computation. Current Biology, 13, 493–497.

Mairal, J., Bach, F., Ponce, J., Sapiro, G., and Zisserman, A. (2009). Supervised dictionary learning. In NIPS’08, pages 1033–1040. NIPS Foundation.

Nair, V. and Hinton, G. E. (2010). Rectified linear units improve restricted boltzmann machines. In Proc. 27th International Conference on Machine Learning.

Olshausen, B. A. and Field, D. J. (1997). Sparse coding with an overcomplete basis set: a strategy employed by V1? Vision Research, 37, 3311–3325.

Pang, B. and Lee, L. (2008). Opinion mining and sentiment analysis. Foundations and Trends in Information Retrieval, 2(1-2), 1–135. Also published as a book. Now Publishers, 2008.

Ranzato, M., Poultney, C., Chopra, S., and LeCun, Y. (2007). Efficient learning of sparse representations with an energy-based model. In NIPS’06.

Ranzato, M., Boureau, Y.-L., and LeCun, Y. (2008). Sparse feature learning for deep belief networks. In NIPS’07, pages 1185–1192, Cambridge, MA. MIT Press.

Salinas, E. and Abbott, L. F. (1996). A model of multiplicative neural responses in parietal cortex. Neurobiol¬ogy, 93, 11956–11961.

Serre, T., Kreiman, G., Kouh, M., Cadieu, C., Knoblich, U., and Poggio, T. (2007). A quantitative theory of immediate visual recognition. Progress in Brain Research, Computational Neuroscience: Theoretical Insights into Brain Function, 165, 33–56.

Snyder, B. and Barzilay, R. (2007). Multiple aspect ranking using the Good Grief algorithm. In Proceedings of HLT-NAACL, pages 300–307.

Vincent, P., Larochelle, H., Bengio, Y., and Manzagol, P.-A. (2008). Extracting and composing robust features with denoising autoencoders. In ICML’08, pages 1096–1103. ACM.

 

9

 

 

 

2020/2021

 

 

1.0 Natural Resources Institute Overview 2

2.0 The Doctoral Thesis 3

3.0 The Doctoral Thesis Process 4

Stage One: Development and Approval of Thesis Research Proposal 5

• Thesis Advisory Committee 5

• Development and Approval of Thesis Research Proposal 6

• Ethics Review Process 8

Stage Two: Candidacy Examination 9

Stage Three: Field and/or Laboratory-Based Research 11

Stage Four: First Draft Meeting 12

• Thesis Format 12

Stage Five: Preparation of Final Thesis 13

Stage Six: Oral Examination 14

Stage Seven: Following Completion of the Oral Examination 19

4.0 Authorship of Publication 20

5.0 Appeal – Procedures and Guidelines 20

6.0 Credential to be Granted 20

 

Ph.D. Thesis Process 2020/2021 - 1 -

 

The degree of Doctor of Philosophy is granted only upon evidence of general proficiency and of distinctive attainment in a special field. In particular, a recognized ability for independent investigation must be demonstrated by the candidate in a thesis which embodies original research or creative scholarship and is presented with a high degree of literacy. It is a research degree and is never conferred solely as a result of coursework study.

The thesis is therefore one of the requirements for the Doctor of Philosophy degree, Natural Resources and Environmental Management. These thesis process guidelines are intended to familiarize Ph.D. students and committee members with the thesis research process. These thesis process guidelines are supplementary to the Faculty of Graduate Studies Thesis Guidelines. 

The Natural Resources Institute (NRI) at the University of Manitoba offers Master's and Ph.D. degrees in resource and environmental management. It was established in 1968 as a degree-granting, interdisciplinary unit with a threefold purpose, namely: (a) to teach management skills leading to a graduate degree of Master of Natural Resources Management (M.N.R.M.); (b) to conduct useful research on actual resource and environmental management problems; and (c) to provide a forum for examining problems in resource use, environmental sustainability, and management.

The Ph.D. program at the NRI was launched in 2001, as part of the Institute’s expanded mandate to contribute to advancing theoretical and applied knowledge in the interdisciplinary fields of resources and the environment. The Ph.D. program prepares students to conduct independent research aimed at solving the complex theoretical and applied issues facing the world environment using a variety of approaches, tools and methods.

 

Ph.D. Thesis Process 2020/2021 - 2 -

 

Ph.D. students in Natural Resources and Environmental Management (NREM) are required to:

• conduct original research;

• synthesize and communicate their findings and conclusions in a thesis; and

• demonstrate a contribution to the theory and practice of natural resources and environmental management.

Through student and faculty research, the Institute is involved in a wide range of natural resource and environmental management issues and problem areas. Research is conducted in conjunction with international agencies, national, provincial and local governments, business and private groups. Sources of the Institute's Ph.D. research support are wide ranging including national and international research funding agencies, client/interest groups, government departments, non-government organizations, and community-based organizations. Students are also financially supported by doctoral scholarships and fellowships from various sources, University of Manitoba Graduate Fellowships, and topical research projects of NRI faculty members.

The research process in general provides a valuable outreach function by bringing together university faculty and practising resource professionals. All thesis research conducted at the Institute is available to the public.

The Ph.D. thesis at the Natural Resources Institute is an independently written research document resulting from an interdisciplinary research project that involves theoretical analysis of a topic or issue, practical application of strategies to address a topic or issue, or a combination of these two, in the areas of natural resources and environmental management. The thesis should show that the student has mastery of the specific field of natural resources and environmental management under investigation and is fully conversant with the relevant literature. The thesis should also demonstrate that the candidate has made a contribution to knowledge in the field of interest. The thesis may entail cooperation with agencies in the community at large and with other faculties at the U of M.

In general, the overall goal of the thesis is to build or apply theory through disciplined and focused independent study. Therefore, the thesis should be based on scholarly study and research that encompasses both theoretical and empirical aspects in the natural resources and environmental management fields, as defined in the Mission  Statement of the Natural Resources Institute.

 

Ph.D. Thesis Process 2020/2021 - 3 -

 

The purpose of the thesis is to:

• develop the student's skills and abilities to conduct independent theoretical and empirical/applied research in the natural resources and environmental management field;

• make original and creative contributions to advancing knowledge; and

• prepare a dissertation demonstrating a high degree of literacy and analytical skills.

As part of the Ph.D. thesis research requirements, each student will:

• prepare a Ph.D. Thesis Research Proposal for committee approval;

• complete a Ph.D. Candidacy Examination;

• complete a first draft of their thesis and have it reviewed by their committee; and

• prepare and defend their Ph.D. Thesis.

The thesis research process has seven distinct stages. Please note that the timing of the seven stages may differ from student to student and that Stages 1 and 2 may occur in reverse order as decided on the advice the Advisor.

 

Each of the preceding stages is described more fully in the material that follows.

Ph.D. Thesis Process 2020/2021 - 4 -

 

Completion of the Ph.D. Thesis Research Seminar; development of the research proposal; Ph.D. Program of Study and Appointment of Advisory Committee; first committee meeting to review and approve the Ph.D. Thesis Proposal  (usually during the first 12-18 months); ethics and animal care  committee approvals; and preparation for the Candidacy Examination.

An Advisory Committee guides the student during the research process. The committee is formally appointed in writing by the NRI Director for each Ph.D. student. The Advisory Committee must consist of a minimum of three members of the Faculty of Graduate Studies, one of whom must hold a primary appointment from within NRI and one of whom must hold a primary appointment outside NRI. Committees may also include one guest member who has expertise in a related discipline/field but is not a member of the Faculty of Graduate Studies. The membership of the Committee, including the advisor, as well as any subsequent changes to it, must be approved by the Dean of the Faculty of Graduate Studies. The Advisor is the Chair of the Advisory Committee.

When the Advisory Committee is formed, the Faculty of Graduate Studies Ph.D.  Program of Study and Appointment of Advisory Committee form and the NRI  Thesis Committee Data form must be completed.

Responsibilities of the committee are to approve the program of study and thesis proposal and to exercise general supervision over the student's work throughout the Ph.D. program. The committee should meet with the student periodically (no less than once a year) to review the student's progress and to report this progress to the Faculty of Graduate Studies on the annual Progress Report form.

The ultimate goal of the committee is to assist the student during the research process. The Institute faculty advisor is formally appointed by the Director based upon the student's stated research interest prior to his/her admission into the program. The faculty advisor provides on-going guidance to the student, directs the work of the committee, and calls and chairs selected meetings of the committee leading to the final oral examination of the thesis. At each stage of the research process, the student must submit draft material to their faculty advisor for review and make necessary changes prior to submission to the committee as a whole. In addition, the faculty advisor and the student are responsible for ensuring that all forms required by the Natural Resources Institute and the Faculty of Graduate Studies are duly and accurately completed.

 

Ph.D. Thesis Process 2020/2021 - 5 -

 

Each student, working in close consultation with their faculty advisor, is responsible for co-ordinating committee meetings and for ensuring that committee members are informed by the NRI, in writing, of meeting date, time and location.

. Students should take into account committee members' schedules in arranging meeting dates. Once the student has identified an appropriate date and time, the faculty advisor and General Office should be notified two weeks in advance using the appropriate form so that proper arrangements for the meeting can be made (confirmation letters and seminar room booking).

If at any time during the research process the student feels a committee member is not fulfilling his/her role, the Director should be approached with the concern. If a student wishes to change faculty advisor at any time, written justification for such a change must be provided to the Director, or Associate Head, if the Director is the faculty advisor.

at the 7000 Level or above,

beyond the Master’s degree (or its equivalent),

A minimum of 6 credit hours of courses must be completed within the Natural Resources Institute and must include NRI 7310 Ph.D. Thesis Research Seminar. The Ph.D. Thesis Research Seminar is offered to assist students with the thesis research and proposal development process.

Changes in program of study, including withdrawal from individual courses or changes of course category, are permitted but only with the approval of the advisor and the Director of the Natural Resources Institute.

Concurrent with the coursework, each student will develop a thesis research proposal in consultation with their faculty advisor and Ph.D. Advisory Committee as the first step in the thesis research process at the NRI.

The proposal should include the following:

• Introduction and Literature Review

• Required subsections:

• Background or Rationale

• Relevant theories

• Management/conservation and/or policy implications

• All of the above must be supported through logical argument and appropriate references to relevant academic literature

• Additional potentially important subsections (to be determined with committee):

 

Ph.D. Thesis Process 2020/2021 - 6 -

 

• Conceptual framework

• Purpose/Goal, Objectives, questions, and/or hypotheses

• Methods/Methodology

• Required subsections:

• Study design

• Sampling methods or Coverage of study subjects

• Analytical methods

• Timeline

• Budget

• Additional potentially important subsections (to be determined with committee):

• Philosophical worldview

• Study area

• Literature/References cited

The advisor will call an Advisory Committee meeting to consider the adequacy of the proposal, at which time a student will give a 20 minute oral presentation and address questions from the committee.

Written notice of the meeting must be provided to each member of the Advisory Committee.

The purpose of the proposal meeting is to:

• enable the Advisory Committee to assess the student’s preparation for undertaking independent Ph.D. research;

• consider the feasibility of the proposed research; and

• provide necessary feedback for further advancement of the research.

The proposal meeting will be chaired by the advisor, and will have the following steps:

1. Introduction by advisor;

2. Presentation of highlights of the research proposal by the student;

3. Review of the proposal by the Advisory Committee; and

4. Determination of approval of the proposal by the Advisory Committee.

The decision of the Advisory Committee must be unanimous. The student is expected to work with the advisor to address any concerns raised in the proposal meeting. The feedback received must be incorporated into the document for circulation among the committee members and the advisor when requested by the committee. The Advisory Committee and NRI Director must approve the oral examination of the thesis proposal on the Faculty of Graduate Studies Ph.D.  Thesis Proposal form, after which the student may proceed to the Candidacy Examination.

 

Ph.D. Thesis Process 2020/2021 - 7 -

 

Completed and approved thesis proposal; development of research design and instruments; and procurement of clearance from the Human Ethics or Animal Care Review Committees.

Ethical conduct related to “human research” activities at the University is governed by The Ethics of Research Involving Human Subjects Policy. All students whose research will involve collecting data from human participants must consult this policy, complete the Course On Research Ethics (CORE), and complete an Ethics Protocol Submission Form. The policy and forms can be obtained from the Office of Research Services, Human Ethics Coordinator, 208-194 Dafoe Road, or by visiting http://umanitoba.ca/research/orec/ethics/humanethicsindex.html.

Submission forms should indicate that the protocol is to be submitted to the Joint-Faculty REB.

The Policy defines as: “any project that involves the collection of specimens, data or information from persons, through intervention or otherwise. Included are procedures that have a low degree of invasiveness (e.g. surveys, interviews, naturalistic observation, exercise or psychometric testing, examination of patient records), as well as more invasive procedures (e.g. blood sampling, insertion of a cannula, administration of a substance).” A

in human research is defined as “a person, who by virtue of his/her involvement in a data-gathering situation or activity, is a source of primary data or information.”

The NRI adheres to the University of Manitoba Policy and Procedures on  Animal Care and Use for research and teaching. All research involving animals must be reviewed and approved by a Protocol Management and  Review Committee prior to being undertaken. It is the responsibility of those needing to use animals to provide sufficient information to the Committee in order that an informed ethical decision can be made regarding their request to use animals. Approval is granted only after the proposed usage and care of the animals, i.e., the protocol of procedures, has been carefully examined. Animal Use Protocol Forms are available online at : http://umanitoba.ca/research/orec/animalcare/animalcareforms.html 

 

Ph.D. Thesis Process 2020/2021 - 8 -

 

The candidacy examination will explore the student’s theoretical and applied knowledge in the broader field of study relating to the dissertation. The focus of the examination will be on the selected thematic areas or subfields (up to four thematic areas), and their synthesis, relating to natural resources and environment and the literature. The candidacy exam is comprised of a written and an oral component. The written examination will normally be completed over a period of three to four weeks. After receiving the Candidacy assignment, the student may seek clarification from members of the examining committee for three days.

(*taken directly from FGS website)

The Candidacy Examination is an absolute requirement of the Faculty of Graduate Studies and, as such, cannot be waived under any circumstances. However, the format and content of the candidacy exam will vary from unit to unit. The purposes of the candidacy exam in doctoral programs is to determine the student's competence in the discipline with respect to understanding and absorbing a broad spectrum of material, and then researching, identifying, analysing, synthesizing, and communicating ideas about that material in depth.

At the time specified by the advisory committee--normally within the first year after the completion of the Ph.D. program coursework but in no case later than one year prior to expected graduation--the student must successfully complete the formal Candidacy Examination.

The examination is conducted according to a procedure established by the unit and approved by the Academic Guide Committee of the Faculty of Graduate Studies. Please see the unit supplemental regulations for the format and composition of the examination committee for the Candidacy Examination. The Candidacy Examination must be held at The University of Manitoba.

This examination, which is independent from the thesis proposal, may be oral, written, or both and may cover subjects relevant to the general area of the candidate's research. These must be made known to the students.

A pass decision of the examiners must be unanimous. Students must be provided with feedback on their performance and access to the reasons for the pass/fail.

 

Ph.D. Thesis Process 2020/2021 - 9 -

 

The Dean of the Faculty of Graduate Studies must be informed whether the candidate has passed or failed the Candidacy Examination on the “Report on Ph.D. Candidacy Examination” form (http://umanitoba.ca/faculties/graduatestudies/forms/index.html).

Any student who fails the Candidacy Examination twice will be required to withdraw from the Faculty of Graduate Studies.

On successful completion of this examination, the student will be considered a candidate for the Ph.D. degree.

Up to four thematic areas or sub-fields relating to the thesis and the broader fields of natural resources and environment will be jointly identified by the thesis Advisory Committee and the student. These thematic areas will serve as the topics of the Candidacy Examination and will encompass the “student’s broader field of study.” The identification of these thematic areas or sub-fields can occur after the Proposal Meeting or if the student and Advisor decide that the Candidacy will occur before the Proposal, the committee and Advisor will need to work together to identify the thematic areas.

The oral examination will normally proceed within one month of the completion of the written component. The oral Candidacy Examination will be conducted by the advisory committee, open to the public, chaired by the Director (or designate), and entail the following:

1. Introduction and explanation of procedures by Chair;

2. Formal presentation by the student for about 25 minutes.

3. Questions by the advisory committee on the selected thematic areas and the student’s broader field of study;

4. In camera determination by advisory committee of pass/fail of student’s Candidacy Examination, both written and oral components; and

5. Completion of Faculty of Graduate Studies form.

Both the written and oral examinations will be evaluated together at the completion of the oral examination. The Chair of the examining committee will call upon each member of the examining committee to determine whether the student has adequate knowledge in the field of study as demonstrated in the written and oral examinations. Considering both the written and oral components of the examination, the committee will vote with regard to the following outcomes:

If this is the first attempt at the Candidacy Examination, the following are potential outcomes of the exam:

• Pass – no further work required – Requires a unanimous decision of the

 

Ph.D. Thesis Process 2020/2021 - 10 -

 

committee.

• Fail - allowed to repeat – Major weaknesses have been identified. After further formal work as determined by the committee, the student will be allowed to re-take the Candidacy Examination. The second attempt of the Candidacy Examinations will normally be completed within six months of the failed attempt.

Note: If allowed to repeat with failure after the first attempt, both the written and oral must be repeated.

If this is the second attempt at the Candidacy Examination, the following are the potential outcomes of the exam:

• Pass – on successful completion of the Candidacy Examination, the student will be considered a candidate for the Ph.D. degree.

• Fail – the Faculty of Graduate Studies will require the student to withdraw from the program.

Collection of primary and secondary data through field and/or laboratory-based research and compilation and processing of the obtained data.

Upon completion of the Candidacy Examination and approval of the research design and instruments (including ethics and animal care approvals), each student is expected to complete field and/or laboratory research work to generate, collect, compile and process primary and secondary data. The nature and duration of field and laboratory work will vary from student to student. Normally, NRI faculty advisors will supervise the student in conducting field research. However, in case of advisor's unavailability, proxy field supervisor should be arranged by the advisor.

Latitude in format to accommodate the subject matter of each thesis is anticipated. Empirical fieldwork and other data collection techniques are critical components of the research process. Accurate records of data sources are essential. Students should ensure that duplicate copies of reports (electronic or hard copy) are safely kept in alternate locations to prevent loss of irreplaceable material. Care should also be taken to maintain confidentiality of material entrusted to the student by agencies or individuals.

 

Ph.D. Thesis Process 2020/2021 - 11 -

 

Presentation and discussion of a first draft of the thesis at an Advisory Committee meeting. The first draft includes the data collected, analysis and discussion of findings in the context of the literature reviewed, and inferences and conclusions drawn.

Prior to the final oral examination of the thesis, the student and the advisor must organize a first draft meeting once a draft of the thesis is complete and deemed ready for presentation to the Advisory Committee for feedback. After data collection and analysis have taken place, writing the first draft of the thesis requires careful presentation of data, analysis and conclusions, and attention to composition, grammar and to the logical organization of the thesis. An appropriate style manual should be consulted prior to thesis writing. An outline is useful in the writing process and care must be taken in the preparation of maps, figures, and tables. The Faculty of Graduate Studies Thesis Guidelines and Information  website outlines the publication requirements of the University.

The student must submit a copy of the first draft of the thesis to each member of the Advisory Committee at least two weeks prior to the meeting.

The Advisory Committee will discuss and recommend modifications necessary for progressing to the final thesis draft and oral examination.

Students should be made aware that the thesis requirement may be met using different formats. Students should discuss possible formats with their faculty advisor and thesis committee members at the proposal meeting.

A typical format often used in the preparation of the thesis is as follows:

Title Page

Signature Page

Abstract

Acknowledgements

Dedication

Table of Contents

List of Tables

List of Figures

Chapter One: Introduction and Statement of the Problem

Chapter Two: Review of Literature

Chapter Three: Methods

Chapter Four: Analysis and Findings

Chapter Five: Discussion and Inferences

 

Ph.D. Thesis Process 2020/2021 - 12 -

 

Chapter Six: Summary, Conclusions, and Recommendations

Footnotes, References, and Appendices

Figures, Illustrations, and Photographs

Additional Materials (i.e. Consent and Access Forms, Copyright Page)

In cases where the student, advisor, and Advisory Committee deem it appropriate, the format for the thesis may take, in part, the form of a manuscript, or manuscripts, ready for submittal to a recognized, peer-reviewed, journal in the appropriate field. The Advisory Committee will ensure that all NRI thesis requirements (i.e., problem or issue statement, objectives, literature review, conclusions, recommendations) are met in the manuscript or chapters as required. The Advisory Committee at the proposal meeting must agree to the journal(s) and manuscript(s) referred to above. The manuscript cannot be in the form of a journal 'note', 'observation', 'opinion', or other 'minor' publication.

Upon completion of the first draft of the thesis, a meeting of the Advisory Committee will be arranged by the advisor to review the thesis for completeness and adequacy to proceed on to the oral defense. The Advisory Committee will advise on the modifications necessary for progressing to the oral defense.

Upon reviewing the first draft document, the Advisory Committee will suggest modifications prior to the final oral examination. Each student is responsible for completing necessary revisions as suggested by committee members. Particular care should be taken to ensure that the linkage between the objectives, hypotheses and/or expected results, findings, conclusions, and recommendations are maintained. In addition, all supporting evidence should be clearly provided throughout the document as appropriate. Figures, tables and maps should be presented in their final form. An abstract and acknowledgements should be included in the final submission. Proper bibliographic referencing should be used according to the appropriate style guide. Detailed publication requirements as described in the Graduate Studies Thesis Guidelines and Information must be followed. The onus is on each student to complete a detailed final editing (grammar, punctuation, spelling and other mechanics) prior to submission to his or her faculty advisor.

Once the thesis is prepared for defence, it is formally submitted by the student to the Faculty of Graduate Studies electronically via the university’s JUMP portal. Detailed instructions for this are provided in a document available at https://umanitoba.ca/faculties/graduatestudies/media/ThesisDistributionPortal.  pdf. Graduate Studies then circulates the thesis to the Advisory Committee members (at this stage called the Examination Committee) for their review. The Examination Committee members must rank the thesis as Class 1 or 2 (see below) in their report to the Dean of Graduate Studies before it is sent to the

Ph.D. Thesis Process 2020/2021 - 13 -

 

External Examiner for their review. The External Examiner must also rank the thesis Class 1 or 2 in order to proceed to the oral examination.

1 The thesis represents a distinct contribution to the candidate’s filed of research and it is acceptable as it stands (or with minor revisions to either content, structure, or writing style) • Thesis has not

received final approval

• Candidate may proceed to their oral examination

2 The thesis has merit since it makes a contribution to the candidate’s field; however, there are research-related concerns that have the potential to be dispelled in the oral examination. The structure and writing are acceptable or require only minor revisions. • Thesis has not

received final approval

• Candidate may proceed to their oral examination

3 The thesis has some merit, but it is not acceptable in the current state since it requires major revisions to one or more of its core components, viz., research content, structure and writing style. • Candidate has

attempt and

to oral

examination

4 The thesis is unacceptable with respect to its core components, viz., research content, structure, and writing style. • Candidate has

attempt and

to oral

examination

The information in this section is adapted from:

http://umanitoba.ca/faculties/graduatestudies/thesis/Ph.D.oegi.html.

Doctoral students must pass an oral examination of the subject of their thesis and matters relating thereto before they may obtain the Ph.D. degree. The examination will be conducted by the Dean of Graduate Studies or the Dean's designate chosen from among scholars of the University. The examining committee will include at least three members of the University of Manitoba academic staff, normally the same members as that of the candidate's advisory committee, and an examiner external to the University of Manitoba. The doctoral candidate, advisor, and all internal members of the examining committee must be present. The presence of the external examiner is strongly encouraged but not mandatory.

 

Ph.D. Thesis Process 2020/2021 - 14 -

 

The following must be assured prior to scheduling the oral examination:

It is required that all internal members of the examining committee be present at the defence, unless exceptional circumstances prevent this and prior approval of the Dean of the Faculty of Graduate Studies has been given. If for any reason an internal examiner is unable to be present in person, participation via video-conferencing may be approved by FGS depending on the constitution of the committee and the presence of other examiners. Oral examinations will be cancelled and deferred if members of the examining committee are not present at the scheduled time and place.

At least two weeks prior to the proposed date the following information must be received by the Faculty of Graduate Studies:

1. Thesis examining committee reports (both internal and external)

2. Student’s examination program information (submitted in a Word document):

a. Biographical data if desired (i.e. birthplace)

b. List of degrees obtained--institution and date

c. List of the candidate's awards, publications, and joint publications

d. An abstract of the thesis (not more than 350 words)

Notes:

1. It is the responsibility of the Department to reserve the room for the oral examination and notify FGS of the date, time and location.

2. The examination must be held at either the University of Manitoba Fort Garry or Bannatyne campus, or the St. Boniface Research Centre.

Once the oral examination has been scheduled, the Faculty of Graduate Studies must be notified. A minimum of two weeks' notice is required. The Faculty of Graduate Studies will then ensure that memoranda announcing the examination are sent to the Department/Unit and the event is posted on the University webpage.

Notes:

1. The oral examination is open to all members of the University community. Graduate students should especially be encouraged to attend.

2. Notices of the examination, including the invitation to graduate students, should be posted within the Department/Unit.

 

Ph.D. Thesis Process 2020/2021 - 15 -

 

Upon receipt of the External Examiner’s report, FGS will forward the report and the “Guidelines for the Scheduling and Conduct of Ph.D. Oral Examinations” to the Department/Unit Head, Advisor, and Internal Examining Committee members.

Approximately one week prior to the oral examination, FGS will provide the Chair of the examination with a Chair package including the following:

• copies of all of the examiner's reports, both internal and external,

• the questions and expected answers, if provided, from the external examiner if not attending,

• the examination programs,

• release and copyright forms for the student,

• "Guidelines for the Scheduling and Conduct of the Oral Examination".

Note:

Questions submitted by the external examiner will not be forwarded to the advisor or to the candidate prior to the examination, unless this is explicitly requested by the external.

The individual chairing the oral examination bears a very important responsibility. A thorough examination and defence of the thesis serves as the University's guarantee that the standards of the Ph.D. are not only maintained, but strengthened. To this end, the procedures for the conduct of the examination must be consistent and adhered to strictly.

The Chair ensures that:

• the location, facilities and equipment are considered appropriate for a Ph.D. defence,

• copies of the examination program are distributed to the candidate, the Examining Committee, and members of the audience, and

• the examining committee is aware of and understands the examination procedures.

The Chair welcomes all of those in attendance and impresses upon the candidate, the examiners, and the audience that this occasion is the senior academic event in the Faculty of Graduate Studies and is a formal examination process prior to awarding the highest academic degree of the University of Manitoba.

 

Ph.D. Thesis Process 2020/2021 - 16 -

 

Referring to the Examination Program, the Chair:

• introduces the candidate, noting previous degrees obtained, awards, special achievements, and publications, and

• introduces the Examining Committee, beginning with a special welcome to the external examiner (if present) and proceeding with introductions of the non-departmental examiner, departmental examiners, and finally the advisor.

Note:

The Chair should note that all members of the examining committee have read the thesis and submitted written reports. The advisor may or may not have included a report.

The Chair outlines the procedures for the conduct of the examination for the candidate, examining committee, and audience as follows:

1. The candidate will be asked to give a 25 to 30 minute presentation of the thesis to include the major objectives, content, results, and conclusions. The Chair also notes that the Candidate's ability to present the thesis clearly and concisely in a 25 to 30 minute period is an essential part of the examination process.

2. At the conclusion of the candidate's presentation, the examiners will be invited to pose questions. The Chair reminds the examiners that they will each be initially allowed approximately 20 minutes to pose their questions and that, because this is an examination on which candidates may, and do, pass or fail, their comments and discussions on the subject matter of the thesis, apart from those essential to the questions, should be made only at the post-examination meeting of the examiners.

3. The order of questioning from the examiners is: external examiner, non-departmental examiner, departmental examiners, and the advisor. The Chair notes that ample time will be given to ensure that the examiners are able to complete their questions and that they may raise additional matters after hearing the initial round of questions. The Chair also notes that, if the external examiner is not present, the Chair (or advisor, as decided) will pose the questions from the external examiner to the candidate.

4. If time permits, and after the questioning by members of the Examining Committee is completed, the Guest member (if applicable and in attendance) may be given an opportunity to ask questions, followed by other members of the FGS and members of the audience.

Ph.D. Thesis Process 2020/2021 - 17 -

 

Notes:

1. The Chair is not permitted to pose any questions of his/her own to the candidate. The Chair must remain neutral at all times.

2. At the conclusion of the candidate's presentation and before the questioning period, the Chair should allow sufficient time for those members of the audience who wish to leave the session to do so.

3. One and one-half hours should normally be sufficient for the conduct of the oral examination by the committee, but in no case should the questioning exceed two hours.

The Chair conducts the remainder of the Oral Examination as follows:

• the audience (and Guest member, if applicable) is thanked for their attendance and support of the candidate and, along with the candidate, is asked to leave so that a confidential meeting of the Examining Committee may be held.

• each examiner is asked, in sequence, to comment on the candidate's performance with respect to the answers to the questions and on whether the candidate should pass or fail the oral examination*.

• the candidate is invited back into the room and informed of the results of the examination, the nature of revisions, and the method of approval of the revisions.

• the candidate is reminded of the procedures for submission of the thesis to MSpace, and of the meaning and use of the copyright and partial release of copyright forms.

• the candidate is congratulated (if appropriate) and members of the Examining Committee thanked.

Notes:

1. *If the external examiner or two or more internal examiners indicate a fail, the candidate fails the examination.

2. The decision of the Examination Committee is reported on the “Final Oral Examination of the Ph.D. Thesis” form, submitted to FGS.

3. A second attempt at the oral defence of the thesis may be scheduled in accordance with the regulations of FGS.

4. The nature of required revisions, prior to submission of the final copy of the thesis to MSpace as well as the procedures for their completion is determined. This is done by the Chair in discussion with the committee.

5. The members of the Examining Committee are asked to sign the approval form, and indicate whether they assign the candidate a vote of “Pass” or “Fail” on their oral defense.

 

Ph.D. Thesis Process 2020/2021 - 18 -

 

• communicate to the Dean of Graduate Studies, through the submission of the “Report for Final Examination of the Ph.D.” form, the nature of the revisions and the procedures to ensure that the revisions are satisfactorily made.

• ensure that the External Examiner is thanked for his/her services.

• approve the revisions to the thesis according to the agreed upon directions from the Examination Committee.

• obtain the committee members’ signatures on the "Final Oral Examination of the Ph.D. Thesis" form, including any members who attended the defence remotely.

• submit the "Final Oral Examination of the Ph.D. Thesis" form to the Graduate Studies Office by the appropriate deadline.

• submit all required forms to the Faculty of Graduate Studies, room 500 University Centre, including thesis release forms and, if applicable, copyright permissions.

• submit a digital version of the final thesis as an e-thesis at the MSpace website(http://mspace.lib.umanitoba.ca).

In addition, the student is recommended for graduation after submitting all the

final copies to the NRI General Office prior to the Faculty of Graduate Studies deadline for submission, required final copies are as follows:

• 1 digital copy – NRI General Office

• 1 digital copy – Advisor

• 1 digital copy – Each Committee member

Students are required to complete all the necessary documentation prior to graduation. Costs for duplicating and binding are the student’s responsibility.

 

Ph.D. Thesis Process 2020/2021 - 19 -

 

Students are encouraged to publish the results of their research work in addition to the preparation of a thesis. The right of first authorship rests with the student for a period of one year from the date of the oral examination of the thesis. Secondary authors should also be agreed upon prior to publication.

After the expiration of the one-year period identified above, or in cases where a thesis is not completed, the faculty advisor and/or other committee members may initiate discussion with the student respecting waiver of publication rights. Any party involved shall refer any dispute respecting publication rights, in the first instance, to the NRI Director or the Associate Head.

Students who disagree with a decision have access to appeal routes and should consult with the Student Advocacy Office, Room 519, University Centre. The Faculty of Graduate Studies Discipline Appeal Form outlines the appeal process and Student Discipline By-Law Appeals Process.

The degree of Doctor of Philosophy, with a specialization in Natural Resources and Environmental Management, is granted upon completion of the requirements of the Natural Resources Institute and the Faculty of Graduate Studies.

 

Ph.D. Thesis Process 2020/2021 - 20 -

 

Twelfth International Multi-Conference on Information Processing-2016 (IMCIP-2016)

Role of Text Pre-Processing in Twitter Sentiment Analysis

Tajinder Singh and Madhu Kumari

National Institute of Technology, Hamirpur 177 005, India

Abstract

Ubiquitous nature of online social media and ever expending usage of short text messages becomes a potential source of crowd wisdom extraction especially in terms of sentiments therefore sentiment classification and analysis is a significant task of current research purview. Major challenge in this area is to tame the data in terms of noise, relevance, emoticons, folksonomies and slangs. This works is an effort to see the effect of pre-processing on twitter data for the fortification of sentiment classification especially in terms of slang word. The proposed method of pre-processing relies on the bindings of slang words on other coexisting words to check the significance and sentiment translation of the slang word. We have used n-gram to find the bindings and conditional random fields to check the significance of slang word. Experiments were carried out to observe the effect of proposed method on sentiment classification which clearly indicates the improvements in accuracy of classification.

© 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license

(http://creativecommons.org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of organizing committee of the Organizing Committee of IMCIP-2016

Keywords: Classification, CRF, n-Gram, Sentiment, Text Pre-Processing.

1. Introduction

Since the early 1990s the use of internet has increased in different forms. People are communicating with each other using various appearances. In the past era the traffic has become almost the double on internet3. With this growth of internet traffic different online social networks such as Facebook, Twitter, LinkedIn, etc are also becoming famous. This in the digital world, things are changing in a very small time and become popular and trendy over OSN (Online Social Network). Different practices of sharing and communicating are not based the content but also on the basis of repetition of the content4. In the recent era micro-blogging has become very common21 and popular platform for all online users. Millions/Billions of users are sharing their opinion on various aspects on very popular and trendy websites such as twitter, Facebook, tumbler, flicker, LinkedIn etc.5 Twitter is a famous micro-blogging and social networking service which provides the facility to users to share, deliver and interpret 140 words’ post known as tweet3, 6. Twitter have 320 M monthly active user. Twitter is accessible through website interface, SMS, or mobile devices. 80% users are active through mobiles7. In the micro-blogging services users make spelling mistakes, and use emoticons for expressing their views and emotions13. Natural language processing is also playing a big role and can be used according to the opinions expressed17.

Corresponding author. Tel.: +91-9882551893.

E-mail address: madhu.jaglan@gmail.com

1877-0509 © 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Peer-review under responsibility of organizing committee of the Organizing Committee of IMCIP-2016

doi:10.1016/j.procs.2016.06.095

 

550 Tajinder Singh and Madhu Kumari / Procedia Computer Science 89 ( 2016 ) 549 – 554

Table 1. Twitter’s User Distribution.

Twitter Distribution Total

Monthly Active users 320 M

Active users on mobile 80%

Language Supported 35

Unique visits monthly to sites with embedded Tweets 1 B

Table 2. Social Text Quality Challenges.

Challenge Description

Stop List Common words frequency of occurrence

Lemmatization Similarity detection of text/words

Text Cleaning Removal of unwanted from the data

Clarity of Words To clear the meaning in text

Tagging Predicting data annotation and its characteristics

Syntax/Grammar Scope of ambiguity, data dependency

Tokenization Various methods to tokenize words or phrases

Representation of Text Various methods and techniques to represent text

Automated Learning Similarity measures and use of characterization

2. Related Work

Due to irregular, short form of text (hlo, whtsgoin etc.), short length and slang text of tweets it is challenging to predict polarity of sentiment text. In sentiment a mixture of applications are needed to study and these all demands large number of sentiments from sentiment holder. A summary of sentiment is needed, as in polarity disambiguation and analysis; a single sentiment is not adequate for decision. A common form of sentiment analysis is aspect based e.g. phone, quality, voice, battery etc.

Rafael Michal Karampatsis8 et al. described the twitter sentiment analysis for specifying the polarity of messages. They used the two stage pipeline approach for analysis. Authors used the sum classifier at each stage and several features like morphological, POS tagging, lexicon etc are identified.

Joao Leal et al.11 worked to classify polarity of messages by using machine learning approaches. Joachim Wagner et al. described work on aspect based polarity classification by using supervised machine learning with Lucie Flekova et al.10 also worked on sentiment polarity prediction in twitter text.

Nathon Aston et al.3 worked on sentiment analysis on OSN. They used a stream algorithm using modified balanced for sentiment analysis. Lifna C.S.4 puts forward a novel approach where the various topics are grouped together into classes and then assign weight age for each class by using sliding window processing model upon twitter streams. In the similar way Emma Haddi et al.12 discussed the role of text pre-processing for sentiment analysis.

Efthymios Kouloumpis14 defined and explained three way sentiment analysis in twitter for identify positive, negative and neutral sentiments. Efstratios Kontopoulos16 proposed a novel approach for analysis of sentiment. The approach is ontology based and it simply find out the sentiment score as well as grade for each distinct notion in the post.

3. Challenges of Social Text Quality

In most of the social media, language used by the users is very informal15. Users create their own words and spelling shortcuts and punctuation, misspellings, slang, new words, URLs, and genre specific terminology and abbreviations. Thus such kind of text demands to be corrected. Thus for analysing the text HTML characters, slang words, emoticons19, stop words, punctuations, urlsetc are needed to be removed. Splitting of attached words are also be noticed for cleansing. Fangxi Zhang et al.9 used Stanford Parser Tools1 for POS tagging and for parsing while the Natural Language Toolkit2 was used for removing stop words and lemmatization.Users who are also rating the product, services and facilities provided by various websites are needed to be addressed. Various systems for analysing users behaviour, views, attitude are needs to be analysed and demands to be normalized. Various shopping and

 

Tajinder Singh and Madhu Kumari / Procedia Computer Science 89 ( 2016 ) 549 – 554 551

customer services supporting websites used various scales like star scale system18 where the highest rating has 5 stars and the lowest rating has only 1 star, binary rating system where 0 and 1 etc. are used which demands to be normalized.

3.1 Text normalization

 

Fig. 1. Text Normalization as Process.

4. Proposed Scheme

The algorithm to deal with slang and identified words of short text massages of twitter used the coexistence of these words with different entities then decides the significance of slang words based on the sentiment strength and probability of co-occurrence of binding words with slang and unidentified words. Various steps involved in the proposed scheme is given below.

 

552 Tajinder Singh and Madhu Kumari / Procedia Computer Science 89 ( 2016 ) 549 – 554 Assumption

Two unidentified words cannot be consecutive in a tweet; binding of this word is spread up to maximum two neighboring words.

Input

Tweet which is having unidentified word, slang word (Ws) and Folksonomies (except emoticons). Output

Insignificance/significance of slang word and if slang is found insignificant then it weeded out from tweet else it is replaced with positive or negative score with respect to hash tag of the tweet.

4.1 Procedure

Step 1: Find the close binding of the slang word with different senses (coexisting) present in collected tweets so far

based on bigram and trigrams language models.

Let Ws be unidentified word and Wxis word sequence which coexist in collected tweets.

Bigram: If we consider bigram language model then mod(Wx) = 1, then this word can occur. Wx Can occur before or after Ws. Collecting prospective sense binding vector C(w), where w is a ordered pair of words and its associated probabilistic weight

C(w) = fall(Wx, P(Wx, P(Wx, Ws) where P(Wx/Ws) or P(Ws /Wx) > 01

P(Wx, Ws) = P(Wx)*P(Ws /Wx) if Wx occurs before Ws.

P(Ws, Wx) = P(Ws)*P(Wx/Ws) if Ws occurs before Wx.

• If WX occurs before and after Ws, then we can use the following equations to resolve this situation: Max(P(Wx, Ws), P(Ws, Wx)

Trigram: If we consider trigram sense binding vector C(w) where w is a ordered triplet of ordered pair of words and their associated probabilistic weight with Ws.

C(w) = fall((Wx 1, Wx2), P(Wx1, Wx2, Ws)) Where P(Wx 1/Ws Wx2), P(Ws /Wx1Wx2) or P(Wx 2/Ws Wx2) > 01

• If Ws occurs within Wx 1 and Wx2 the following combination:

Ws Wx 1Wx2, WsWx2Wx1, Wx 1Ws Wx2, Wx2Ws Wx1, Wx 1Wx2Ws and Wx2Wx 1Ws,

• Then

P(Wx1, Wx2, Ws) = max(P(Ws Wx 1Wx2), P(WsWx2Wx 1), P(Wx 1WsWx2), P(Wx 2WsWx 1), P(Wx 1Wx 2Ws), P(Wx 2Wx 1Ws)).

At this stage no filtering is done. We try to collect possible bindings.

Step 2: Analysis of these binding of slang word based on fields associated with coexisting words using Conditional Random Fields (CRF) is done at this stage to decide the significance of Ws. Using CRF Part of Speech (POS) tagging of the tweet which contains Ws is done then significance of Ws the measured using following rules:

 

Tajinder Singh and Madhu Kumari / Procedia Computer Science 89 ( 2016 ) 549 – 554 553

 

 

 

Fig. 2. Distribution of Sentiments over Dataset before and after the Pre-processing.

 

Fig. 3. Accuracy of Proposed Scheme under Different Variants of Preprocessing.

 

a. If Ws occurs before and after a proper noun then it is significant.

b. If is coexisting with collective noun and has reference to a proper noun then Ws is less significant.

c. Else Ws is insignificant and Ws can be weeded out of tweet.

Step 3: After collecting the significant all Ws these words are replaced by positive/negative sentiment scores with respect to the concept present in the binding set computed in step1. Following procedure is used to compute the sentiment score (SentiScore) of Ws.

SentiScore(Ws)  max (P(Ws, Wx)  senti(Wx))

P(Ws, Wx) is computed in first step and Wx is a vector, where senti(Wx) max[t (xi)]. t (xi) is sentiment of the tweet in which xi word which is component of Wx is present sentiment of tweet which has Ws is updated as follows:

t(Ws)told  SentiScore(Ws)

told is earlier sentiment value of tweet which was holding Ws.

5. Experiments and Results

For experimentation we have used twitter corpus data. More description of data can be found in2, this data comprises of six fields, first field is sentiment class of the tweet which are negative, neutral and positive, represented by 0, 2 and 4 respectively, rest of field are the id of the tweet, the date of the tweet, the query, the user that tweeted, the text of the tweet. In order to evaluate and measure the impact of proposed scheme on the sentiment classification task we have used Support Vector Machine (SVM) based classifier. We carried out experiment in to two phases, in the first phase we applied the proposed scheme of normalization to the tweets’ text by ignoring their sentiment class. After the normalization process we consider the sentiment class and class 2 i.e. is resolved in to new classes as: 1 (less negative) and 3 (less positive) based on the sentiments of unidentified (slang) words.

Results of experiments clearly suggest that proposed scheme not only robust to size of data but also perform better in terms of accuracy of sentiment classification.

 

554 Tajinder Singh and Madhu Kumari / Procedia Computer Science 89 ( 2016 ) 549 – 554

6. Conclusions and Future Scope

This work is to analyse the impact of pre-processing and normalization on short massages like tweets which are full of information, noise, symbols, abbreviations, folksonomy and unidentified words. Looking at the interestingness to interpret the slang and unidentified words in tweets towards the sentiment, this paper focuses to identify the importance of slang words and to measure their impact on sentiment of the tweet. The proposed scheme used in this paper first gathers the coexisting words with the slang and then exploits characteristics of these binding words to define the significance and sentiment strength of slang word used in the tweet which not only facilitate the better sentiment classification but also ensure the sturdiness of classier as shown in the results. It is yet to be seen the how well the proposed scheme will perform with different classifiers on text streams.

References

[1] http://cs.stanford.edu/people/alecmgo/trainingandtestdata.zip

[2] Twitter Sentiment Classification using Distant Supervision

[3] N. Aston, T. Munson, J. Liddle, G. Hartshaw, D. Livingston and W. Hu, Sentiment Analysis on the Social Networks Using Stream Algorithms, Journal of Data Analysis and Information Processing, vol. 2, pp. 60–66, (2014).

[4] C. S. Lifna and M. Vijayalakshmi, Identifying Concept-Drift in Twitter Streams, ICACTA-2015, Elsevier, (2015).

[5] Ayushi Dalmia, Manwitter Sentiment Analysis Thish Gupta, Vasudeva Varma, The Good, the Bad, and the Neutral, Sem Eval (2015).

[6] Santhi Chinthala, Ramesh Mande, Suneetha Manne and Sindhura Vemuri, Sentiment Analysis on Twitter Streaming Data, Springer International Publishing Switzerland, (2015).

[7] http://twittercommunity.com

[8] Rafeal Mcheal Karampatsis, John Pavlopoulos and Prodromos Malakasiotis, Sentiment Analysis Two Stage Sentiment Analysis of Social Network Messages, SemEval, (2014).

[9] Fangxi Zhang, Zhihua Zhang and Man Lan, ECNU: A Combination Method and Multiple Features for Aspect Extraction and Sentiment Polarity Classification, Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), Dublin, Ireland, pp. 252–258, 23–24 August (2014).

[10] Lucie Flekov, Oliver Ferschk and Iryna Gurevych, UKPDIPF: A Lexical Semantic Approach to Sentiment Polarity Prediction in Twitter Data, Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), Dublin, Ireland, pp. 704–710, 23–24 August (2014).

[11] Joao Leal, Sara Pinto, Ana Bento and Hugo Gonc¸alo Oliveira, Paulo Gomes, CISUC-KIS: Tackling Message Polarity Classification with a Large and Diverse set of Features, Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), Dublin, Ireland, pp. 166–170, 23–24 August (2014).

[12] Emma Haddi, Xiaohui Liu and Yong Shi, The Role of Text Pre-processing in Sentiment Analysis, First International Conference on Information Technology and Quantitative Management, Elsevier, (2013).

[13] Apoorv Agarwal, Boyi Xie, Ilia Vovsha, Owen Rambow and Rebecca Passonneau, Sentiment Analysis of Twitter Data, Department of Computer Science Columbia University New York, NY 10027 USA.

[14] Efthymios Kouloumpis, Theresa Wilson and Johanna Moore, Twitter Sentiment Analysis: The Good the Bad and the OMG!, Proceedings of the Fifth International AAAI Conference on Weblogs and Social Media, (2011).

[15] Sara Rosenthal, Alan Ritter, Preslav Nakov and Veselin Stoyanov, SemEval-2014 Task 9: Sentiment Analysis in Twitter, Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), Dublin, Ireland, pp. 73–80, 23–24 August (2014).

[16] Efstratios Kontopoulos, Christos Berberidis, Theologos Dergiades and Nick Bassiliades, Ontology Based Sentiment Analysis of Twitter Posts, Expert Systems with Applications, vol. 40, issue 10, pp. 4065–4074, August (2013).

[17] Chetashri Bhadane, Hardi Dalal and Heenal Doshi, Sentiment Analysis-Measuring Opinions, International Conference on Advanced Computing Technologies and Applications (ICACTA), vol. 45, pp. 808–814, (2015).

[18] Xing Fang and Justin Zhan, Sentiment Analysis Using Product Review Data, Journal of Big Data, 2015 Springer, (2015).

[19] Xia Hu, Jiliang Tang, Huiji Gao and Huan, Liu, Unsupervised Sentiment Analysis with Emotional Signals, Proceedings of the 22nd International Conference on World Wide Web, WWW’13, ACM, (2013).

[20] Ana Mihanovi´c, Hrvoje Gabelica and ˇZivko Krsti, Big Data and Sentiment Analysis using KNIME: Online Reviews vs. Social Media, MIPRO 2014, 26–30 May 2014, Opatija, Croatia, pp. 1463–1468, (2014).

[21] Lowri Williams, Christian Bannister, Michael Arribas-Ayllon, Alun Preece and Irena Spasic, The Role of Idioms in Sentiment Analysis, Expert Systems with Applications, Elsevier, (2015).

[22] http://www.iprospect.com/en/ca/blog/10-sentiment-analysis-tools-track-social-marketing-success/

[23] http://marcobonzanini.com/2015/03/02/mining-twitter-data-with-python-part-1/

 

NMED-T: A TEMPO-FOCUSED DATASET OF CORTICAL AND

BEHAVIORAL RESPONSES TO NATURALISTIC MUSIC

Steven Losorelli1,2 Duc T. Nguyen1,2 Jacek P. Dmochowski3 Blair Kaneshiro1,2

1Center for the Study of Language and Information, Stanford University, USA

2Center for Computer Research in Music and Acoustics, Stanford University, USA

3Department of Biomedical Engineering, City College of New York, USA

{slosorelli, dtn006, blairbo}@stanford.edu jdmochowski@ccny.cuny.edu

 

ABSTRACT

Understanding human perception of music is foundational to many research topics in Music Information Retrieval (MIR). While the field of MIR has shown a rising interest in the study of brain responses, access to data remains an obstacle. Here we introduce the Naturalistic Music EEG Dataset—Tempo (NMED-T), an open dataset of electro-physiological and behavioral responses collected from 20 participants who heard a set of 10 commercially available musical works. Song stimuli span various genres and tem¬pos, and all contain electronically produced beats in du¬ple meter. Preprocessed and aggregated responses include dense-array EEG and sensorimotor synchronization (tap¬ping) responses, behavioral ratings of the songs, and basic demographic information. These data, along with illustra-tive analysis code, are published in Matlab format. Raw EEG and tapping data are also made available. In this pa-per we describe the construction of the dataset, present re¬sults from illustrative analyses, and document the format and attributes of the published data. This dataset facilitates reproducible research in neuroscience and cognitive MIR, and points to several possible avenues for future studies on human processing of naturalistic music.

1. INTRODUCTION

Humans possess a unique ability to process music, and many topics in Music Information Retrieval (MIR) involve computational modeling of human perception. Tasks that humans often perform with ease—such as melody extrac¬tion, beat detection, and artist identification—remain open topics in MIR. At the same time, a full understanding of the cognitive and perceptual processes underlying human processing of music has yet to be reached.

Greater cross-disciplinary collaboration between MIR and neuroscience has been proposed [14], and a number of studies have incorporated approaches from both fields.

c Steven Losorelli, Duc T. Nguyen, Jacek P. Dmochowski, and Blair Kaneshiro. Licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0). Attribution: Steven Losorelli, Duc T. Nguyen, Jacek P. Dmochowski, and Blair Kaneshiro. “NMED-T: A Tempo-Focused Dataset of Cortical and Behavioral Responses to Nat¬uralistic Music”, 18th International Society for Music Information Re¬trieval Conference, Suzhou, China, 2017.

 

For example, neural correlates of short- and long-term fea¬tures introduced in MIR for genre classification [34] have been sought [1, 6, 10, 20], and brain responses have been used in MIR-related applications including tempo estima¬tion [29,30] and emotion recognition [5, 21]. Yet even as brain data become more prevalent in MIR research, ex¬perimental design, data collection, and data cleaning can present challenges [14]. Therefore, the research com¬munity can arguably benefit from curated, ready-to-use datasets of brain responses to real-world musical works.

Aiming to provide an open dataset with which sev¬eral MIR and neuroscience topics can be explored, we introduce the Naturalistic Music EEG Dataset—Tempo (NMED-T), a dataset of EEG and behavioral responses to commercially available musical works. The dataset con¬tains dense-array EEG responses from 20 adult participants who listened to 10 full-length songs, as well as tapped re¬sponses to the beat of shorter excerpts (collected in a sep¬arate listen). These responses have been cleaned and ag¬gregated, and are ready to use in Matlab format along with ratings of familiarity and enjoyment, as well as basic de¬mographic information about the participants.

NMED-T contributes to a growing body of publicly available music-related EEG repositories, including the DEAP [17], Music BCI [32], NMED-H [15], and Open-MIIR [31] datasets. It is well suited for MIR research in that the data are cleaned and ready to use but are also made available in raw form; stimuli are complete, naturalistic 1 musical works spanning a wide range of tempos; metadata links to stimulus audio are provided; and behavioral data are included. Moreover, as EEG was recorded while par¬ticipants listened attentively but did not focus on any par¬ticular dimension of the songs, these data are suitable for studying many aspects of music processing.

The remainder of the paper is structured as follows. In  2 we describe stimulus selection, study design, data col¬lection, and data preprocessing. Illustrative analyses of the preprocessed data, which build upon past music perception and MIR approaches and reveal cortical and behavioral ev¬idence of entrainment to musical beat, are presented in  3. In  4 we document the dataset itself. We conclude and discuss potential future uses of the data in  5.

1 Denoting real-world music—i.e., music that was created to be con¬sumed in everyday life, as opposed to controlled stimuli created for ex¬perimental research.

 

339

 

340 Proceedings of the 18th ISMIR Conference, Suzhou, China, October 23-27, 2017

 

2. METHODS

2.1 Stimuli

2.1.1 Stimulus Selection

As the present dataset is focused on naturalistic music and tempo, stimuli were ecologically valid, real-world musical works containing steady, electronically produced beats in duple meter at a variety of tempos. The 10 selected songs are all 4:30–5:00 in length, contain vocals (all but one in English), and are in the Western musical tradition. Song information is summarized in Table 1.

To aid in song selection, we computed objective mea-sures of tempo using publicly available Matlab code [8]. The computed tempos were then validated perceptually by four trained musicians. The final set of selected songs range in tempo from 56–150 BPM—a wide enough range to potentially explore octave errors [11, 35]. To facilitate further research on the audio as well as the responses, we purchased digital versions of all songs from Amazon, and include in Table 1 each song’s Amazon Standard Identifi¬cation Number (ASIN).

These real-world stimuli are complex and contain en-ergy at various frequencies—not just those directly re-lated to the beat. We followed the approach of Nozaradan et al. [27] and visualized low-frequency spectra of the stim¬uli. We extracted the amplitude envelope of each song us¬ing the MIR Toolbox, version 1.5 [18] at a sampling rate of 125 Hz (the sampling rate of the preprocessed EEG), and plotted magnitude spectra up to 15 Hz. As can be seen in Fig. 1, spectral peaks often occur at harmonics and subhar-monics of the beat—implicating the hierarchical timescale of music—as well as at other frequencies.

2.1.2 Stimulus Preparation

To prepare the stimuli for the EEG experiment, full-length songs were first converted to mono using Audacity, ver¬sion 2.1.2.2 We then embedded the second audio channel with an intermittent click that was transmitted directly to the EEG amplifier (not played to participants) to ensure precise time stamping of the stimuli. For the behavioral experiment, we created 35-second excerpts of each song. Using Audacity, we selected the audio from 1:00–1:34 and applied a linear fade-in and fade-out to the first and last 2 seconds, respectively. We then appended 1 second of si¬lence to make the conclusion of each excerpt more obvious to the participant.

2.2 Participants

Twenty right-handed participants, aged 18–29 years (mean age 23 years, 6 female) participated in the experiment. All reported normal hearing, fluency in English, and no cogni¬tive or decisional impairments. We imposed no eligibility criteria related to formal musical training; 17 participants reported having received training (mean 8.4 years among those with training). Participants reported listening to mu¬sic for 14.5 hours per week on average.

2http://www.audacityteam.org

 

2.3 Experimental Specifications & Data Collection

This study was approved by the Stanford University In-stitutional Review Board. All participants provided writ-ten informed consent before participating. Each partici-pant filled out a general demographic and musical back-ground questionnaire, after which the EEG and tapping blocks were completed, with the EEG block always oc-curring first.

2.3.1 EEG Experiment

First, each participant was informed that the general pur-pose of the experiment was to study human processing of music, and that he or she would be completing an EEG ses¬sion and a behavioral test. As the EEG data were collected for the general study of music processing (not limited to beat perception), no explicit mention of beat or tempo was given at this stage of the experiment. Rather, participants were instructed simply to listen attentively to the songs as they played, and to avoid movement of any kind (includ¬ing stretching, yawning, and tapping or moving to the beat) during the trials. Songs were presented in random order. Following each trial, participants delivered ratings of fa¬miliarity and enjoyment for the song just presented, on a scale of 1–9. The EEG experiment was split into two con¬secutive recording blocks in order to mitigate participant fatigue, limit data size of the EEG recordings, and allow for verification of electrode impedances between record¬ings. Therefore, a total of 40 EEG recordings were col¬lected across the 20 participants.

The EEG experiment was programmed in Matlab ver-sion 2013b 3 with a custom template built on the Psy-chophysics Toolbox, version 3 [4]. Each participant sat comfortably in a chair at a desk for the duration of the experiment. Stimuli were presented through magnetically shielded Genelec 1030A speakers at a measured loudness level between 73–78 dB. During the trials, the participant viewed a fixation image presented on a computer monitor located 57 cm in front of him or her.

Dense-array EEG was recorded using the Electrical Geodesics, Inc. (EGI) GES300 system [33]. Data were recorded from 128 electrodes with vertex reference using an EGI Net Amps 300 amplifier and Net Station 4.5.7 acquisition software, sampled at 1 kHz with a range of 24 bits. Electrode impedances were verified to be no greater than 50 kΩ—an appropriate level for this system— at the start of each recording.

2.3.2 Behavioral Experiment

Following the EEG recordings, the electrode net was re-moved from the participant, and the behavioral test began. Here, each participant listened to the 35-second song ex¬cerpts, after receiving instructions to “tap to the steady beat of the song as you perceive it.” If the participant had ques¬tions about tapping to multiple tempos for a given song, he or she was instructed to tap to the steady beat that best re¬flected his or her perception of it in the moment. Excerpts were presented in random order.

3https://www.mathworks.com

 

Proceedings of the 18th ISMIR Conference, Suzhou, China, October 23-27, 2017 341

# Song Title Artist ASIN Tempo (BPM) Tempo (Hz) min:sec

1 “First Fires” Bonobo B00CJE73J6 55.97 0.9328 4:38

2 “Oino” LA Priest B00T4NHS2W 69.44 1.1574 4:31

3 “Tiptoes” Daedelus B011SAZRLC 74.26 1.2376 4:36

4 “Careless Love” Croquet Club B06X9736NJ 82.42 1.3736 4:54

5 “Lebanese Blonde” Thievery Corporation B000SF16MI 91.46 1.5244 4:49

6 “Canop´ee” Polo & Pan B01GOL4IB0 96.15 1.6026 4:36

7 “Doing Yoga” Kazy Lambist B01JDDVIQ4 108.70 1.8116 4:52

8 “Until the Sun Needs to Rise” R¨uf¨us du Sol B01APT6JKA 120.00 2.0000 4:52

9 “Silent Shout” The Knife B00IMN40O4 128.21 2.1368 4:54

10 “The Last Thing You Should Do” David Bowie B018GS2A46 150.00 2.5000 4:58

Table 1. Stimulus set. Songs were selected on the basis of vocals, electronically produced beats, genre, tempo, and length.

 

Figure 1. Low-frequency magnitude spectra of stimulus amplitude envelopes. Frequencies related to the musical beat hierarchy, from 1/4x the tempo (whole notes) to 8x the tempo (32nd notes) are denoted with vertical dashed lines.

 

Tapping responses were collected using Tap-It, an iOS application that plays audio while simultaneously record-ing responses tapped on the touchscreen [16]. We note a tap-to-timestamp latency of approximately 15 msec (st. dev. 5 msec) [16]. An Apple iPad 2 was used for this experiment, with stimuli delivered at a comfortable listen-ing level using over-ear Sony MDR-V6 headphones.

2.4 Data Preprocessing

All data preprocessing and analysis was conducted using Matlab, versions 2013b and 2016b.

2.4.1 EEG Preprocessing

The following preprocessing steps were performed on in-dividual EEG recordings that had been exported from Net Station to Matlab cell arrays. First, data from each electrode in the electrodes-by-time data matrix were zero-phase filtered using 8th-order Butterworth highpass (0.3 Hz) and notch (59–61 Hz) filters, and a 16th-order Chebyshev Type I lowpass (50 Hz) filter. Following this, the filtered data were temporally downsampled by a factor of 8 to a final sampling rate of 125 Hz.

We extracted trial labels, onsets, and behavioral rat-ings, and corrected the stimulus onset times using the click events sent directly from the audio to the EEG amplifier. The data for each trial were epoched, concatenated, and DC corrected (subtracting from each electrode its median value). Bad electrodes were removed from the data ma¬trix, resulting in a reduction in the number of rows. We computed EOG components for tracking vertical and hori¬zontal eye movements, and retained electrodes 1–124 for further analysis, excluding electrodes on the face. We 

 

applied a validated approach using Independent Compo-nents Analysis (ICA) to remove ocular and cardiac artifacts from the data [2,13] using the runica function from the EEGLAB toolbox [7].

As final preprocessing steps, transients exceeding 4 standard deviations of each electrode’s mean power were identified in an iterative fashion and replaced with NaNs. We then reconstituted missing rows corresponding to pre-viously identified bad electrodes with rows of NaNs, en-suring that each data matrix contained the same number of rows. We appended a row of zeros—representing the vertex reference—and converted the data frame to average reference (subtracting from each electrode the mean of all electrodes). All missing values (NaNs) were imputed with the spatial average of data from neighboring electrodes, and a final DC correction was performed. Finally, the epochs were separated once again into single trials. There¬fore, after preprocessing, each recording produced a cell array of EEG data, each element of which contained an electrodes-by-time matrix of size 125  T, where T varied according to the length of the stimulus.

After preprocessing all recordings, we aggregated the data on a per-song basis. The data frame for each song is thus a 3D electrodes-by-time-by-participant matrix of size 125  T  20.

2.4.2 Preprocessing of Tapping Responses

The Tap-It application stores the timestamps of taps, in seconds, measured from the device touchscreen on a per-trial basis, as well as each participant’s randomized stim-ulus ordering array [16]. We aggregated the tapping re-sponses in a cell array and the ordering arrays in a matrix.

 

342 Proceedings of the 18th ISMIR Conference, Suzhou, China, October 23-27, 2017

 

3. ILLUSTRATIVE ANALYSES

The following analyses are presented to illustrate basic properties of the dataset.

3.1 EEG Responses

One approach to studying beat processing using EEG in-volves low-frequency ( 20 Hz) steady-state evoked po-tentials (SS-EPs). In an SS-EP paradigm, stimuli presented (e.g., flashed or sounded) at a particular frequency elicit brain responses at that same frequency. While SS-EPs are more often used to study vision processing [25], the ap¬proach has in recent years been used to study responses to auditory rhythms. Here, SS-EPs have shown evidence of entrainment to musical beat, peaking at beat- and meter-related frequencies even when metrical accents are imag¬ined [26] or when beat frequencies do not dominate low-frequency stimulus spectra [27]. To our knowledge, mu¬sic SS-EP studies have to date used simple, synthesized rhythmic patterns as stimuli. Our first illustrative analysis extends this approach to complex, naturalistic music.

Spatial filtering is a technique for EEG analysis whereby a weighted sum of electrodes is computed sub-ject to some criterion [3]. Advantages of concentrating ac¬tivity of interest from many electrodes to a few spatially filtered components include dimensionality reduction, im¬proved SNR, and a reduction in multiple comparisons. For the present analysis we consider two simple spatial filters. The first is simply the mean across all electrodes (ME), which can be thought of as a constant weight applied to each electrode. For the second, we perform Principal Com¬ponents Analysis (PCA), and analyze the first PC of data.

We first averaged each song’s 3D electrodes-by-time-by-participant matrix across participants, producing an electrodes-by-time matrix for each song. Then, so that we analyzed the same amount of data for each song and to account for the time course of listener entrainment to the beat [9], we retained 4 minutes of data from each song, starting 15 seconds into the song.

To compute the spatial filters, we concatenated the participant-averaged data frames across all songs, produc-ing an electrodes-by-aggregated-time matrix. Then, for the ME spatial filter, we computed the mean across electrodes, while for the PCA filter we computed electrode weightings for PC1 using Singular Value Decomposition (SVD). Fi¬nally, we reshaped each resulting song-concatenated com¬ponent vector into a songs-by-time matrix. As our current interest is on SS-EPs, we present the magnitude spectrum of each component on a per-song basis.

The SS-EPs are shown in Fig. 2; y-axis scaling is con-sistent within each spatial filtering technique. By inspec-tion of the plots, low frequencies (<15 Hz) of ME spectra occasionally contain peaks at frequencies in the musical beat hierarchy (e.g., Song 5). PC1 performs better, elic-iting more robust spectral peaks at beat-related frequen-cies. Moreover, EEG PC1 appears to peak at frequencies directly related to musical beat, while suppressing many of the other spectral peaks that were observed in the magni¬tude spectra of stimulus amplitude envelopes (Fig. 1).

 

Spatial filters can be visualized by projecting the filter weights on a 2D scalp topography. While it is common to convert the spatial filter weights to a so-called “forward model,” which captures the projection of filtered activity on the scalp, for PCA the spatial filter is equivalent to the forward model [28]. The ME filter, applying a constant weight to all electrodes, would reveal no spatial variation. However, the PC1 filter topography (Fig. 2, bottom right) applies a range of positive and negative weights to the elec¬trodes, which may help to explain why this filter produces more prominent spectral peaks at beat frequencies.

3.2 Behavioral Ratings

Participant ratings of familiarity and enjoyment are shown in Fig. 3. Familiarity with the songs was low overall; rat¬ings of enjoyment tended to be higher, and also varied more across participants.

3.3 Tapped Responses

For each trial of tapping data, we first converted each inter-tap interval to an instantaneous measure of tempo in Hz, mapped it to the midpoint of the interval, and then linearly interpolated the result to a consistent timing grid with a temporal resolution of 2 Hz. We analyze and plot data from a 17-second interval starting 15 seconds into the excerpt (i.e., starting at time 1:15 in the complete song).

The aggregate tapping responses are shown in Fig. 4. We present two visualizations of these results. First, the top figure for each song shows instantaneous tempo over the time of the excerpt for individual participants (gray curves), with the median across participants plotted in black. In bottom figures, we compute the median tempo across time for each individual participant, and summa-rize with histograms. Beat-related frequencies are shown in the orange (1/2x tempo frequency), green (tempo fre-quency), and red (2x tempo frequency) lines. To a large extent, participants tended to tap at what we had previously determined to be the tempo frequency. However, there are cases of lower agreement, particularly for the slowest songs (Song 1 and Song 2). Here, the histograms suggest a nearly bimodal distribution of tapped tempos, split be¬tween the computational measure and twice that, with the higher measure lying closer to what is considered the pre¬ferred tempo region for humans [23].

4. PUBLISHED DATASET

We publish the cleaned EEG data, aggregated behavioral ratings, aggregated tapped responses, and basic demo-graphic data about the participants in Matlab .mat for-mat. Example code and helper functions for the illustrative analyses are provided, also in Matlab format. Finally, we publish raw EEG recordings (for researchers who wish to apply their own preprocessing pipelines) as well as individ¬ual files of the tapped responses. The dataset is available for download from the Stanford Digital Repository [22], 4 published under a Creative Commons CC-BY license.

4https://purl.stanford.edu/jn859kj8079

 

Proceedings of the 18th ISMIR Conference, Suzhou, China, October 23-27, 2017 343

 

 

Figure 2. Low-frequency EEG spectra using a mean-electrode spatial filter (top) and PC1 spatial filter (bottom) for each song. Beat-related frequencies are shown with dashed vertical lines. Bottom right: PC1 spatial filter weights.

 

 

Figure 3. Participant ratings of familiarity and enjoyment.

4.1 Cleaned EEG Data

The .mat file songSS Imputed.mat contains the cleaned EEG records, aggregated across participants, for song SS (§ 2.4.1). There are 10 such files, one per song. Each .mat file contains the following variables:

• dataSS: 3D electrodes-by-time-by-participant data frame. The size is 125 × T × 20, with T varying according to the song.

• subsSS: Cell array of participant ids. Contents are the same for all songs, but are included in order to link these data to raw EEG files, raw tapping re-sponses, and participant demographics.

• fs: Sampling rate, in Hz (always 125).

4.2 Raw EEG Data

We provide the raw EEG records in their exported state before preprocessing. No filtering, epoching, or cleaning has been performed. As each participant underwent two recordings, there are a total of 40 raw EEG files. The file PP R raw.mat refers to recording R 1, 2 from partici¬pant PP. Each file contains the following variables:

• X: Raw data frame. Size is electrodes-by-time, 129× T, where T is the total length of the recording, including time periods not related to the experimen¬tal trials. The vertex reference electrode is row 129.

 

• DIN 1: Cell array containing all event labels (trig-gers) and times. We provide the helper function parseDIN.m to extract the labels and onsets into numeric vectors. Full specification on labels is pro¬vided in the README file accompanying the dataset.

• fs: Sampling rate, in Hz (always 1000).

4.3 Behavioral Ratings

Participants delivered ratings of familiarity (Q1) and en-joyment (Q2) of each song during the EEG session. The file behavioralRatings.mat contains a single vari¬able behavioralRatings, which is a 3D participant-by-song-by-question (20 × 10 × 2) matrix.

4.4 Tapping Responses

Aggregated and raw tapping responses are stored in the file TapIt.zip. This archive contains the file TapIt.mat, which comprises the following variables:

• allTappedResponses: Aggregated tapped response times across all participants and songs. This is a participants-by-song (20 × 10) cell array. Each entry is a column vector of tap times in seconds, recorded from the device touchscreen.

• allSongOrders: Song-order vectors, aggregated across all participants. This is a participants-by-trial (20×10) matrix, where each row contains the stimu¬lus presentation order for the respective participant. Numbering starts at 1.

Individual response files are also included in the .zip file:

• PPP SS.txt: Single trial of tapped responses, in sec-onds, for participant PPP and song SS.

• PPP play order.txt: Stimulus presentation ordering for participant PPP. Numbering starts at 0.

4.5 Participant Demographics

The file participantInfo.mat contains a struct ar¬ray participantInfo with participant demographics. Fields

 

344 Proceedings of the 18th ISMIR Conference, Suzhou, China, October 23-27, 2017

 

 

Figure 4. Tapping responses. Top: Instantaneous tempo over time for individual participants (gray), with median across participants in black. Bottom: Histograms of median tempo, over time, for individual participants. Ground-truth tempos are shown with orange (1/2x tempo frequency), green (tempo frequency), and red (2x tempo frequency) lines.

 

include age, nYearsTraining, weeklyListening (hours), and id (participant identifier link to raw filenames).

4.6 Code

The file Code.zip contains the Matlab scripts for the analyses performed in  3. A variety of helper functions and files (e.g., electrode location map, script to parse the DIN 1 variable in raw EEG files) are also provided here.

5. DISCUSSION

This paper introduces NMED-T, an open dataset of elec-trophysiological and behavioral responses collected from 20 participants listening to real-world musical excerpts. The published data include both raw and preprocessed dense-array EEG and tapping responses, behavioral ratings of the songs, and basic demographic information.

Our illustrative analyses validate the frequency-tagging, SS-EP approach [26,27] with responses to complex, nat¬uralistic music (Fig. 2). Even a simple PCA filter com¬puted from trial-averaged responses highlights beat-related frequencies in the EEG spectra. Many PC1 spectra show prominent peaks between 5–10 Hz, regardless of tempo; future research could use this dataset to investigate further the stimulus and response attributes contributing to this phenomenon. The variability in tapping responses (Fig. 4) highlights the challenge of defining a ‘ground truth’ for tempo and beat identification, particularly for complex mu¬sic [24]. Here we see various, sometimes conflicting re¬sults across and within participants’ tapped responses. Past research has suggested that humans inherently prefer cer¬tain frequencies related to natural movement [23, 35]. This may help to explain why some participants tapped at twice the tempo for the slowest songs, tending toward the postu¬lated 2-Hz natural resonant frequency.

We faced several trade-offs when designing the study. Collection of EEG data, while relatively inexpensive [14], still incurs costs of equipment and time. Participant fa-tigue must also be taken into account when planning the overall duration of an experiment. As we wished to col-lect EEG responses to a set of full-length songs from ev-ery participant, we were limited in the number of songs 

 

we could use, and relegated the secondary tapping task to shorter excerpts. Stimulus selection, too, is often a com-promise of breadth and depth. For example, the OpenMIIR dataset [31] uses shorter stimuli from a variety of genres, but at the expense of depth within any one genre; while the NMED-H [15] includes various stimulus manipulations of complete songs, but only four songs from a single genre. Our focus on full-length songs with a steady beat and a va¬riety of tempos limited the range of genres somewhat. We also deliberately avoided massively popular songs in or¬der to minimize possible effects, on the brain responses, of varying familiarity, established personal preferences, and autobiographical associations with the songs [12].

There are shortfalls to the dataset. One potential con-found is that the EEG session always preceded the behav¬ioral task; thus, participants were more familiar with the music during the tapping task. As a result, the tapping data may not be suitable for studying the time course of beat entrainment. However, we chose this arrangement so that participants would not be focused specifically on beat while EEG responses were recorded. Second, the tapping data show variations in tapped tempo across participants and within-participant over time. Whether this reflects our participant pool (not all trained musicians), inadequate in¬struction for the task, or is merely characteristic of this re¬sponse is not addressed in the present illustrative analyses. Finally, listeners are known to exhibit variations in tempo octave during tapping while largely agreeing on whether a song is fast or slow [19], but we unfortunately did not collect data here to explore this distinction.

Generally speaking, this dataset facilitates research on encoding and decoding of naturalistic music. While the study design and initial analyses focused primarily on beat and tempo, the EEG responses can be analyzed in conjunc¬tion with various other stimulus features as well. Investiga¬tion of individual differences is also possible (e.g., predict¬ing a particular participant’s tapping tempo or preference rating from his or her own EEG). Other researchers might consider augmenting the dataset with complementary re¬sponses to the same songs. Ideally, the dataset will find ap¬plications in MIR and neuroscience research beyond those envisioned by the authors of this study.

 

Proceedings of the 18th ISMIR Conference, Suzhou, China, October 23-27, 2017 345

 

6. ACKNOWLEDGMENTS

This research was supported by the Patrick Suppes Gift Fund and the Wallenberg Network Initiative: Culture, Brain, Learning. The authors thank Anthony Norcia, Dan Ellis, and the anonymous ISMIR reviewers for helpful feedback on the manuscript.

7. REFERENCES

[1] V. Alluri, P. Toiviainen, I. P. J¨a¨askel¨ainen, E. Glerean,

M. Sams, and E. Brattico. Large-scale brain networks emerge from dynamic processing of musical timbre, key and rhythm. NeuroImage, 59(4):3677–3689, 2012.

[2] A. J. Bell and T. J. Sejnowski. An information-maximization approach to blind separation and blind deconvolution. Neural Computation, 7(6):1129–1159, 1995.

[3] B. Blankertz, R. Tomioka, S. Lemm, M. Kawanabe, and K. R. Muller. Optimizing spatial filters for ro¬bust EEG single-trial analysis. IEEE Signal Processing Magazine, 25(1):41–56, 2008.

[4] D. H. Brainard. The psychophysics toolbox. Spatial Vi¬sion, 10(4):433–436, 1997.

[5] R. Cabredo, R. S. Legaspi, P. S. Inventado, and M. Nu-mao. An emotion model for music using brain waves. In ISMIR, pages 265–270, 2012.

[6] F. Cong, V. Alluri, A. K. Nandi, P. Toiviainen, R. Fa, B. Abu-Jamous, L. Gong, B. G. W. Craenen, H. Poiko-nen, M. Huotilainen, and T. Ristaniemi. Linking brain responses to naturalistic music through analysis of on¬going EEG and stimulus features. IEEE Trans. Multi¬media, 15(5):1060–1069, 2013.

[7] A. Delorme and S. Makeig. EEGLAB: An open source toolbox for analysis of single-trial EEG dynamics in¬cluding independent component analysis. Journal of Neuroscience Methods, 134(1):9–21, 2004.

[8] D. P. W. Ellis. Beat tracking by dynamic programming. Journal of New Music Research, 36(1):51–60, 2007.

[9] P. Fraisse and B. H. Repp. Anticipation of rhythmic stimuli: Speed of establishment and precision of syn-chronization (1966). Psychomusicology: Music, Mind, and Brain, 22(1):84, 2012.

[10] N. Gang, B. Kaneshiro, J. Berger, and J. P. Dmo-chowski. Decoding neurally relevant musical features using Canonical Correlation Analysis. In ISMIR, 2017.

[11] J. Hockman and I. Fujinaga. Fast vs slow: Learning tempo octaves from user data. In ISMIR, pages 231– 236, 2010.

[12] P. Janata. The neural architecture of music-evoked autobiographical memories. Cerebral Cortex, 19(11):2579–2594, 2009.

[13] T.-P. Jung, C. Humphries, T.-W. Lee, S. Makeig, M. J. McKeown, V. Iragui, and T. J. Sejnowski. Extended ICA removes artifacts from electroencephalographic recordings. In NIPS, pages 894–900, 1998.

 

[14] B. Kaneshiro and J. P. Dmochowski. Neuroimaging methods for music information retrieval: Current find¬ings and future prospects. In ISMIR, pages 538–544, 2015.

[15] B. Kaneshiro, D. T. Nguyen, J. P. Dmochowski, A. M. Norcia, and J. Berger. Naturalistic music EEG dataset—Hindi (NMED-H). In Stanford Digital Repos¬itory, 2016.

[16] H.-S. Kim, B. Kaneshiro, and J. Berger. Tap-It: An iOS app for sensori-motor synchronization experiments. In ICMPC12, 2012.

[17] S. Koelstra, C. Muhl, M. Soleymani, J. S. Lee, A. Yaz-dani, T. Ebrahimi, T. Pun, A. Nijholt, and I. Patras. DEAP: A database for emotion analysis using phys¬iological signals. IEEE Trans. Affective Computing, 3(1):18–31, 2012.

[18] O. Lartillot and P. Toiviainen. A Matlab toolbox for musical feature extraction from audio. In DAFx, pages 237–244, 2007.

[19] M. Levy. Improving perceptual tempo estimation with crowd-sourced annotations. In ISMIR, pages 317–322, 2011.

[20] Y. P. Lin, J. R. Duann, W. Feng, J. H. Chen, and T. P. Jung. Revealing spatio-spectral electroencephalo-graphic dynamics of musical mode and tempo percep¬tion by independent component analysis. Journal of NeuroEngineering and Rehabilitation, 11(1), 2014.

[21] Y. P. Lin, C. H. Wang, T. P. Jung, T. L. Wu, S. K. Jeng, J. R. Duann, and J. H. Chen. EEG-based emo-tion recognition in music listening. IEEE Transactions on Biomedical Engineering, 57(7):1798–1806, 2010.

[22] S. Losorelli, D. T. Nguyen, J. P. Dmochowski, and B. Kaneshiro. Naturalistic music EEG dataset— Tempo (NMED-T). In Stanford Digital Repository, 2017.

[23] D. Moelants. Preferred tempo reconsidered. In ICMPC7, pages 1–4, 2002.

[24] D. Moelants and M. F. McKinney. Tempo perception and musical content: What makes a piece fast, slow, or temporally ambiguous? In ICMPC8, pages 558–562, 2004.

[25] A. M. Norcia, L. G. Appelbaum, J. M. Ales, B. R. Cot-tereau, and B. Rossion. The steady-state visual evoked potential in vision research: A review. Journal of Vi¬sion, 15(6):4, 2015.

[26] S. Nozaradan, I. Peretz, M. Missal, and A. Mouraux. Tagging the neuronal entrainment to beat and meter. The Journal of Neuroscience, 31(28):10234–10240, 2011.

[27] S. Nozaradan, I. Peretz, and A. Mouraux. Selective neuronal entrainment to the beat and meter embed-ded in a musical rhythm. The Journal of Neuroscience, 32(49):17572–17581, 2012.

 

346 Proceedings of the 18th ISMIR Conference, Suzhou, China, October 23-27, 2017

[28] L. C. Parra, C. D. Spence, A. D. Gerson, and P. Sajda. Recipes for the linear analysis of EEG. NeuroImage, 28(2):326–341, 2005.

[29] A. Sternin, S. Stober, J. A. Grahn, and A. M. Owen. Tempo estimation from the EEG signal during percep¬tion and imagination of music. In BCMI/CMMR, 2015.

[30] S. Stober, T. Pr¨atzlich, and M. Meinard. Brain beats: Tempo extraction from EEG data. In ISMIR, pages 276–282, 2016.

[31] S. Stober, A. Sternin, A. M. Owen, and J. A. Grahn. Towards music imagery information retrieval: Intro-ducing the OpenMIIR dataset of EEG recordings from music perception and imagination. In ISMIR, 2015.

[32] M. S. Treder, H. Purwins, D. Miklody, I. Sturm, and B. Blankertz. Decoding auditory attention to instruments in polyphonic music using single-trial EEG classification. Journal of Neural Engineering, 11(2):026009, 2014.

[33] D. M. Tucker. Spatial sampling of head electrical fields: The geodesic sensor net. Electroencephalogra-phy and Clinical Neurophysiol., 87(3):154–163, 1993.

[34] G. Tzanetakis and P. Cook. Musical genre classifica-tion of audio signals. IEEE Trans. Speech and Audio Processing, 10(5):293–302, 2002.

[35] L. van Noorden and D. Moelants. Resonance in the per¬ception of musical pulse. Journal of New Music Re¬search, 28(1):43–66, 1999.

 

Curriculum Vitae

&

Faculty Personal Record

Inas S. Khayal

Assistant Professor

The Dartmouth Institute of

Health Policy & Clinical Practice

Geisel School of Medicine at Dartmouth

Adjunct Assistant Professor

Department of Computer Science

Dartmouth College

 

Modified: August 11, 2017

46 Centerra Pkwy, Rm 256

Lebanon, NH 03766

Email: Inas.Khayal@dartmouth.edu

Of~ice Phone: (603) 653-6866

Website: http://sustainablehealth.dartmouth.edu

 

(This page is intentionally left blank.)

 

Contents

1 Education 2

1.1 Degreed Programs  2

1.2 Professional Development & Training 3

2 Professional Experience 4

2.1 Academic Appointments  4

2.2 Relevant Research & Development Experience  5

2.3 Consulting Experience  6

2.4 Entreprenurial Experience  6

2.5 Honors and Awards  7

3 Research 8

3.1 Publications, Written Works & Presentations  8

4 Teaching 20

4.1 Academic Supervision 20

4.1.1 Summary 20

4.1.2 Other Advising: Research Projects & Internships 20

4.1.3 Postdoctoral Advising 21

4.2 Curriculum Development 21

4.2.1 New Syllabus & Course Development 21

4.3 Teaching Experience 21

4.3.1 Faculty Teaching Experience 21

5 Academic & Professional Service 22

5.1 Academic Service 22

5.1.1 Program & Departmental Service 22

5.1.2 Institutional Service 22

5.2 Professional Service 23

5.2.1 Professional Committees 23

5.2.2 Funding Review 23

5.2.3 Journal Review 23

5.2.4 Conference Review 23

5.2.5 Conference Organization 23

 

1

 

1 Education

1.1 Degreed Programs

Sep 2009 University of California, Berkeley & San Francisco: San Francisco, CA

Ph.D., Joint Bioengineering Doctoral Program

Developed techniques for acquisition, reconstruction, and quantitative analysis of Dif-fusion Tensor Imaging for brain tumor diagnosis throughout the diseased state (i.e. de¬termined tumor type, determined the most pathologic region to be surgically removed, evaluate the effectiveness of drug therapies, etc.) and designed clinical studies to test these developments.

Seven journal papers published, 12 oral and poster conference presentations.

Dec 2008 University of California, Berkeley: Berkeley, CA

2008 Berkeley Mayfield Fellows Program, The Walter A. Haas School of Business – a uni¬versity program that offers 10 selected students per year in-depth training and experi¬ence in high-tech entrepreneurship. It also includes a summer internship at a venture capital firm, a VC mentor, and a study abroad experience.

Dec 2007 University of California, Berkeley: Berkeley, CA

Management of Technology Program, The Walter A. Haas School of Business, the College of Engineering and the School of Information

May 2004 Boston University: Boston, MA

Bachelor of Science, Biomedical Engineering, Summa Cum Laude

Dean’s List (all semesters)

Most Outstanding Senior Project in Biomedical Engineering for ”Angular Dependence

of Light Scattering By Apoptotic Cells”

 

2

 

1. EDUCATION

1.2 Professional Development & Training

2011 MIT Course MAS 961 Networks, Complexity and Its Applications

2010 New England Complex Systems Institute

CX102: Computer Programming and Complex Systems

CX201: Complex Physical, Biological, and Social Systems

CX202: Complex Systems Modeling and Networks

2010 MIT Course ESD 15.871 Introduction to Systems Dynamics

2010 MIT Course ESD 69 Seminar on Healthcare Systems Innovation

2010 MIT Course ESD 34 Systems Architecture

2008 California Institute for Quantitative Biosciences

QB3 Global Bio-Entrepreneurship Course

 

3

 

2 Professional Experience

2.1 Academic Appointments

July 2016 - Dartmouth College, Department of Computer Science Hanover, NH

Present Adjunct Assistant Professor of Computer Science

Developed & teaching CS 89/189 Health Informatics and advise students.

Oct 2015 - Dartmouth College, Geisel School of Medicine Hanover, NH

Present Assistant Professor of Health Policy & Clinical Practice

The Dartmouth Institute for Health Policy & Clinical Practice

Leads the Sustainable Health Lab: Develops systems solutions to curb the growth of chronic disease by improving health outcomes, minimizing healthcare costs and as-sessing technical and social biomedical innovations.

Dartmouth Hitchcock Collaboratory for Population Health Faculty Member: Preven-tion and management of complex conditions in primary care through innovative mod-els of care, uses of technology and new measures and predictive analytics.

Awarded as a Robert Wood Johnson Foundation 2017 New Connections Scholar. Awarded as an AcademyHealth 2017 Systems Science Scholar - to stimulate systems ap¬proaches to complex social, behavioral, and environmental influences.

July 2010 - Masdar Institute of Science & Technology Abu Dhabi, UAE

Aug 2015 Assistant Professor of Engineering Systems & Management

Leads the Sustainable Health Lab: Working with Stakeholders, designing human sub-jects research studies and building quantitative models based on Biological, Social and Environmental Factors affecting Non-communicable Chronic Diseases.

July 2011- Massachusetts Institute of Technology Cambridge, MA

Aug 2015 Research Affiliate at The Media Lab

Collaborating on mobile monitoring technologies to measure social and biological fac 

tors affecting health state with Professors Alex (Sandy) Pentland and Kent Larson.

July 2010 - Massachusetts Institute of Technology Cambridge, MA

June 2011 Visiting Scholar at The Media Lab

Collaborating on mobile monitoring technologies to measure social and biological fac 

tors affecting health state with Professors Alex (Sandy) Pentland and Kent Larson.

 

4

 

2. PROFESSIONAL EXPERIENCE

2.2 Relevant Research & Development Experience

Sep 2004 - University of California, SF, Surbeck Lab of Advanced Imaging San Francisco, CA

Sep 2009 Graduate Researcher

Developed techniques for acquisition, reconstruction, and quantitative analysis of Dif-fusion Tensor Imaging for brain tumor diagnosis throughout the diseased state (i.e. de¬termined tumor type, determined the most pathologic region to be surgically removed, evaluatet the effectiveness of drug therapies, etc.) and designed clinical studies to test these developments.

Sep 2003 - Boston University, Biomedical Optics and Biophotonics Laboratory Boston, MA

May 2004 Independent Student Researcher, Undergraduate Bioengineering Senior Thesis Designed and constructed an instrument to measure light scattering from cells as a function of angle.

Data analysis and modeling to assess changes in the cell nuclei of apoptotic cells. Project includes written proposal, formal technical research report, and a research presentation conference. Program included experience in patent procedures, federal regulations, project planning and quality control.

Summer GE Medical Systems, Early Identification Program Milwaukee, WI

2003 Intern, CT Advanced Technology and Applications (included Six Sigma Training), Designed a tissue phantom and experiment to evaluate the impact of human tissue elasticity on Image Quality for various CT acquisition parameters. Required Advanced Workstation and Animation Shop.

Determined the effect of multiple post-processing filters on 16-slice cardiac image data using Minitab.

Developed a web-based database of human biophysical data and anatomical specifica-tions to aid in design of emerging CT Technologies.

Designed a study to investigate the impact of patient breath-hold and motion affects on CT image quality in the lung. The results were presented at the Conference on Screen¬ing for Lung Cancer mentioned above.

Summer Boston Scientific Corporation Natick, MA

2002 Intern, Molecular Interventions

Developed injectable polymeric drug delivery systems for delivery with minimally in-vasive medical devices. Used spectrofluorometric methods to quantitate DNA release kinetics from polymer systems. Submitted four invention disclosures on novel injec-tion catheter designs; Disclosure received patent approval.

Intern, Urology

Optimized injectable formulations for urological applications. Performed viscometry measurements on these formulations, evaluated the systems in bench-top tissue mod-els and participated in acute animal studies (prepared materials, recorded data and summarized results).

Summer Boston University, Hearing Research Center Boston, MA

2001 Independent Student Researcher

Designed an adaptive test program to measure just noticeable differences (JNDs) us-ing MATLAB interfaced with TDT System II that created a signal according to the speci¬fied frequency. The project determined the correlation between interference and hand dominance with JND values.

 

5

 

2. PROFESSIONAL EXPERIENCE

Jun 1999 - ImmunoGen, Inc. Cambridge, MA

Dec 2000 Independent Researcher and Intern

Performed a toxicity test for various carcinogens on the growth of breast and skin can¬cer cells.

Awarded Bridgewater Regional Science Fair- First Place 2000, Massachusetts State Sci¬ence Fair- Third Place 2000, Naval Academy Award, and $1,500 Sea Education Associ¬ation Award for an ImmunoGen sponsored science fair project. Skills gained: microp-pipetting, ELISA, gel electrophoresis and Western Blot.

Oct 1998 - Boston University, Biology Department Boston, MA May 1999 Independent Researcher

Determined the effects of ultraviolet radiation on Escherichia coli’s RecA and UvrA pathways. Biology Lab sponsored science fair project; Awarded Bridgewater Regional Science Fair Second Place 1999 and Massachusetts State Science Fair Third Place 2000.

2.3 Consulting Experience

Oct 2008 - QB3 New Ventures Consulting (QNVC) San Francisco, CA

Aug 2009 Consultant,

Served as a consultant in the QNVC team to explore market opportunities for scientific

platforms developed at UCSF.

Jun 2008 - 360Fresh, Inc. San Francisco, CA May 2009 Project Manager Consultant,

Developed data mining and analysis tools. Researched and compiled Market Research reports

Jun 2007 - Ortho Kinematics, Inc. San Francisco, CA

May 2008 Clinical Analyst Consultant,

Performed clinical assessment and clinical competitive analysis

Sep 2002 - Boston University Information Technology, ResNet Boston, MA May 2004 Student Consultant,

Diagnosed networking problems and determined the proper resolutions to customers’ computer issues.

2.4 Entreprenurial Experience

Jan 2013 - Engineering Systems Analytics LLC Mansfield, CT

Present Founder, CFO, & COO

A family-owned engineering consultancy startup that applies control, automation, sig¬nal processing and information technologies to the domains of smart power grids, wa 

ter distribution, transportation, manufacturing, and sustainable health.

 

6

 

2. PROFESSIONAL EXPERIENCE

2.5 Honors and Awards

2017 New Connections Scholar Award Robert Wood Johnson Foundation

2017 Systems Science Scholars Award Academy Health

2017 Best Paper Award 19th International Conference on Bioinformatics, Com¬putational Biology and Biomedical Engineering

2013 Honorable Mention Paper Award 7th International Conference on Axiomatic Design

2008 Mayfield Fellow, Management of Technology Program University of California, Berkeley

2004-

2007 Graduate Opportunity Fellowship University of California, Berkeley

2004 Most Outstanding Senior Project Award Boston University

 

7

 

3 Research

3.1 Publications, Written Works & Presentations

Summary

Patents: [5, 4, 3, 2,1]

Peer-Reviewed (Journal) Articles: [18,17, 16, 15, 14, 13, 12, 11, 10, 9, 8, 7, Peer-Reviewed (Conference) Articles: [17,16, 15, 14, 13, 12, 11, 10, 9,8,7, Reviewed Abstracts: [24,23, 22, 21, 20, 19, 18, 17, 16, 15, 14, 13, 12, 11, 10, Op-Eds and Media: [4, 3, 2,1]

Presentations (International): [24,23, 22, 21, 20, 19, 18, 17, 16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2,1]

Presentations (National): [24,23, 22, 21, 20, 19, 18, 17, 16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1]

Presentations (Regional/Local): [9, 8, 7, 6, 5, 4, 3, 2,1]

Theses: [2, 1]

White Papers: [5,4, 3, 2, 1]

Students for which Prof. Inas Khayal was the advisor or co-advisor are highlighted in forest green. Prof. Khayal’s role in the publication is highlighted in blue.

Presentations where Prof. Inas Khayal was extended an invitation to present are indicated with an (*). Presentations where Prof. Inas Khayal presented a poster/talk are indicated with an (#).

Patents

[5] I. Khayal, W. Naimark, and M. Palasis, “Multiple needle injection catheter and method of use of same,” United States Patent US7 442187 B2, October 28, 2008.

[4] I. Khayal, W. Naimark, and M. Palasis, “Multiple needle injection catheter and method of use of same,” European Patent EP 1841486, October 10, 2007.

[3] T. Freyman, I. Khayal, and W. Naimark, “Apparatus and methods for delivering a bolus of thera-peutic material,” United States Patent US 20 060 217 656, September 28, 2006.

[2] T. Freyman, I. Khayal, and W. Naimark, “Apparatus and methods for delivering a bolus of thera-peutic material,” International Patent WO/2006/098 816, September 21, 2006.

 

8

 

PEER-REVIEWED (JOURNAL) ARTICLES

[1] I. Khayal, W. Naimark, and M. Palasis, “Multiple needle injection catheter and method of use of same,” International Patent WO/2006/081 026, March 08, 2006.

Peer-Reviewed (Journal) Articles

[18] I. Khayal and A. Farid, “A Dynamic System Model for Personalized Healthcare Delivery and Man-aged Individual Health Outcomes,” Submitted to Enterprise Information Systems, 2017.

[17] I. Khayal and A. Farid, “Architecting a System Model for Personalized Healthcare Delivery and Managed Individual Health Outcomes,” Complexity (in press), 2017.

[16] I. Khayal, W. Zhou, and J. Skinner, “Structuring and Visualizing Healthcare Claims Data Using Systems Architecture,” International Journal of Biological, Biomolecular, Agricultural, Food and Biotechnological Engineering, vol. 11, no. 4, pp. 342–346, 2017.

[15] I. S. Khayal and A. M. Farid, “The Need for Systems Tools in the Practice of Clinical Medicine,” Systems Engineering, vol. 20, no. 1, pp. 3–20, Jan 2017. [Online]. Available: http://dx.doi.org/10.1002/sys.21374

[14] I. Khayal and A. M. Farid, “Axiomatic Design Based Human Resources Management for the En-terprise Transformation of the Abu Dhabi Healthcare Labor Pool,” Journal of Enterprise Transfor-mation, vol. 5, no. 3, pp. 162–191, 2015.

[13] M. Butt, S. Quan, A. S. Pentland, and I. Khayal, “Gender Differences in Real-Home Sleep of Young and Older Couples,” Southwest Journal of Pulmonary and Critical Care, vol. 10, no. 1, pp. 289–299, 2015.

[12] M. Butt, T. B. M. J. Ouarda, S. F. Quan, A. S. Pentland, and I. Khayal, “Technologically sensed social exposure related to slow-wave sleep in healthy adults,” Sleep and Breathing, no. 19, pp. 255–261, 2015.

[11] T. Ouarda, C. Charron, K. N. Kumar, P. R. Marpu, H. Ghedira, A. Molini, and I. Khayal, “Evolution of the rainfall regime in the United Arab Emirates,” Journal of Hydrology, vol. 514, pp. 258–270, 2014.

[10] R. O. Oloritun, T. B. Ouarda, S. Moturu, A. Madan, A. S. Pentland, and I. Khayal, “Change in BMI Accurately Predicted by Social Exposure to Acquaintances,” PLOS One, vol. 8, no. 11, p. e79238, 2013.

[9] R. Oloritun, A. Madan, A. Pentland, and I. Khayal, “Identifying Close friendships in a Sensed So-cial Network,” Procedia-Social and Behavioral Sciences, vol. 79, pp. 18–26, 2013.

[8] N. Aharony, W. Pan, C. Ip, I. Khayal, and A. Pentland, “Social fMRI: Investigating and shaping social mechanisms in the real world,” Pervasive and Mobile Computing, vol. 7, no. 6, pp. 643–659, 2011. [Online]. Available: http://linkinghub.elsevier.com/retrieve/pii/S1574119211001246

[7] I. S. Khayal, S. R. Vandenberg, K. J. Smith, C. P. Cloyd, S. M. Chang, S. Cha, S. J. Nelson, and T. R. McKnight, “MRI apparent diffusion coefficient reflects histopathologic subtype, axonal disrup-tion, and tumor fraction in diffuse-type grade II gliomas.” Neuro-oncology, vol. 13, no. 11, pp. 1192–1201, 2011.

 

9

 

PEER-REVIEWED (CONFERENCE) ARTICLES

[6] I. S. Khayal, M.-Y. C. Polley, L. Jalbert, A. Elkhaled, S. M. Chang, S. Cha, N. A. Butowski, and S. J. Nelson, “Evaluation of diffusion parameters as early biomarkers of disease progression in glioblastoma multiforme,” Neurooncology, vol. 12, no. 9, pp. 908–916, 2010. [Online]. Available: http://www.ncbi.nlm.nih.gov/pubmed/20501631

[5] I. S. Khayal and S. J. Nelson, “Characterization of low-grade gliomas using RGB color maps derived from ADC histograms,” Journal of Magnetic Resonance Imaging, vol. 30, no. 1, pp. 209–213, Jul 2009. [Online]. Available: http://dx.doi.org/10.1002/jmri.21810

[4] I. S. Khayal, T. R. McKnight, C. McGue, S. Vandenberg, K. R. Lamborn, S. M. Chang, S. Cha, and S. J. Nelson, “Apparent diffusion coef~icient and fractional anisotropy of newly diagnosed grade II gliomas,” NMR in Biomedicine, vol. 22, no. 4, pp. 449–455, may 2009. [Online]. Available: http://dx.doi.org/10.1002/nbm.1357

[3] W. Bian, I. S. Khayal, J. M. Lupo, C. Mcgue, S. Vandenberg, K. R. Lamborn, and S. M. Chang, “Mul-tiparametric Characterization of Grade 2 Glioma Subtypes Using Magnetic Resonance Spectro-scopic, Perfusion and Diffusion Imaging,” Translational Oncology, vol. 2, no. 4, pp. 271–280, 2009.

[2] F. Crawford, I. Khayal, C. McGue, S. Saraswathy, A. Pirzkall, S. Cha, K. Lamborn, S. Chang, M. Berger, and S. Nelson, “Relationship of pre-surgery metabolic and physiological MR imaging parameters to survival for patients with untreated GBM,” Journal of Neuro-Oncology, vol. 91, no. 3, pp. 337–351, 2009. [Online]. Available: http://dx.doi.org/10.1007/s11060-008-9719-x

[1] I. S. Khayal, F. W. Crawford, S. Saraswathy, K. R. Lamborn, S. M. Chang, S. Cha, T. R. McKnight, and S. J. Nelson, “Relationship between choline and apparent diffusion coefficient in patients with gliomas,” Journal of Magnetic Resonance Imaging, vol. 27, no. 4, pp. 718–725, apr 2008. [Online]. Available: http://dx.doi.org/10.1002/jmri.21288

Peer-Reviewed (Conference) Articles

[17] I. Khayal and A. Farid, “A Dynamic Model for a Cyber-Physcal Healthcare Delivery System with Human Agents,” in Proceedings of the 2017 IEEE International Conference on Systems, Man, and Cybernetics (SMC2017), Intelligent Industrial System Special Issue, October 2017.

[16] I. Khayal and A. Farid, “Designing Smart Cities for Citizen Health & Well-being,” in Proceedings of the 2017 IEEE International Summer School on Smart Cities (IEEE S3C), August 2017.

[15] I. Khayal and A. Farid, “An Architecture for a Cyber-Physical Healthcare Delivery System with Human Agents,” in Proceedings of the 2017 IEEE International Summer School on Smart Cities (IEEE S3C), August 2017.

[14] I. Khayal, M. McGovern, M. Bruce, and S. Bartels, “Developing an Integrated Behavioral Health System Model using Engineering Design,” in Proceedings of the 2017 Institute of Industrial and Systems Engineering Annual Conference, Pittsburgh, PA, May 2017.

[13] I. Khayal, W. Zhou, and J. Skinner, “Structuring and Visualizing Healthcare Claims Data Using Sys¬tems Architecture,” in Proceedings of the 19th International Conference on Bioinformatics, Compu¬tational Biology and Biomedical Engineering, Boston, MA, April 2017.

[12] I. Khayal and A. M. Farid, “The Application of Model-Based Systems Engineering to the Practice of Clinical Medicine,” in Proceedings of the 2017 Annual IEEE International Systems Conference (SysCon), Montreal, Quebec, Canada, April 2017, pp. 1–6.

10

 

PEER-REVIEWED (CONFERENCE) ARTICLES

[11] A. Popoola, M. Butt, T. B. M. J. Ouarda, and I. Khayal, “Automatically Capturing Sleep and Social Factors to Understand Ramadan in the Real World,” in IEEE-EMBS International Conference on Biomedical and Health Informatics (BHI), Valencia, Spain, jun 2014.

[10] R. Oloritun, A. Pentland, and I. Khayal, “Dynamics of Human Social Networks: People, Time, Relationships, and Places,” in Collaborative Innovation Networks Conference (COINs13). Chile: eprint arXiv:1308.1287, aug 2013. [Online]. Available: http://adsabs.harvard.edu/abs/ 2013arXiv1308.1287O

[9] I. S. Khayal and A. M. Farid, “Axiomatic Design Based Volatility Assessment of the Abu Dhabi Healthcare Labor Market: Part II - Case Study,” in The Seventh International Conference on Ax-iomatic Design (ICAD2013), Worcester, MA, USA, jun 2013, pp. 77–83.

[8] A. M. Farid and I. S. Khayal, “Axiomatic Design Based Volatility Assessment of the Abu Dhabi Healthcare Labor Market: Part I - Theory,” in The Seventh International Conference on Axiomatic Design (ICAD2013), Worcester, MA, USA, jun 2013, pp. 71–76.

[7] M. Butt, S. T. Moturu, A. S. Pentland, and I. Khayal, “Automatically captured sociability and sleep quality in healthy adults,” in 35th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC’13). Osaka, Japan: Ieee, jul 2013, pp. 4662–4665. [Online]. Available: http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=6610587

[6] R. Oloritun, A. Madan, A. Pentland, and I. Khayal, “Evolution of Social Encounters in Ad-Hoc Mobile Face-to-Face Interaction Networks,” in 2012 ASE International Conference on Social Informatics (SocialInformatics), vol. 0, Alexandria, VA, USA, dec 2012, pp. 192–198. [Online]. Available: http://doi.ieeecomputersociety.org/10.1109/SocialInformatics.2012.101

[5] R. O. Oloritun, A. Pentland, A. Madan, and I. Khayal, “Identifying Close friends interactions in Face-to-Face Networks,” in Applications of Social Network Analysis (ASNA2012), Zurich, Switzer-land, Sep 2012, pp. 1–19.

[4] N. Aharony, W. Pan, C. Ip, I. Khayal, and A. Pentland, “The Social fMRI: Measuring, Understanding, and Designing Social Mechanisms in the Real World,” in 13th International Conference on Ubiq¬uitous Computing (UbiComp 2011). Beijing, China: Massachusetts Institute of Technology, Sep 2011, pp. 445–454.

[3] S. T. Moturu, I. Khayal, N. Aharony, W. Pan, and A. S. Pentland, “Using Social Sensing to Understand the Links between Sleep, Mood, and Sociability,” in 2011 IEEE Third Intl Conference on Privacy Security Risk and Trust and 2011 IEEE Third Intl Conference on Social Computing, no. SocialCom. Boston, MA, USA: IEEE, Sep 2011, pp. 208–214. [Online]. Available: http://web.media.mit.edu/{~}panwei/pub/mood{_}socialcom2011.pdf

[2] S. T. Moturu, I. Khayal, N. Aharony, W. Pan, and A. S. Pentland, “Sleep, mood and sociability in a healthy population,” in 2011 Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBS’11), vol. 2011, Media Lab, Massachusetts Institute of Technology, Cambridge, MA 02139 USA. Boston, MA, USA: IEEE, Sep 2011, pp. 5267–5270. [Online]. Available: http://www.ncbi.nlm.nih.gov/pubmed/22255526

[1] J. Sairamesh, K. Argenbright, H. C. Simmons, R. Nemana, R. Rajagopal, I. Khayal, and P. Anderson, “Early Warning Methods for Improving Patient Care: Harnessing Insight from Raw Patient Notes,” in 3rd INFORMS Workshop on Data Mining and Health Informatics

 

11

 

REVIEWED ABSTRACTS

(DM-HI 2008), vol. 12, no. 10, Washington, D.C., USA, oct 2008, p. s146. [Online]. Available: http://www.ncbi.nlm.nih.gov/pubmed/24091256

Reviewed Abstracts

[24] I. Khayal and A. Farid, “The Application of Systems Tools in the Practice of Clinical Medicine,” in Proceedings of the 5th International Engineering Systems Symposium (CESUN2016), Washington, D.C., June 2016.

[23] I. Khayal and A. Farid, “Developing Sustainable Health Frameworks in Smart Cities,” in Proceed-ings of the First IEEE International Smart Cities Conference, Guadalajara, Mexico, September 27, 2015.

[22] M. Butt, S. Moturu, A. Pentland, and I. Khayal, “Deep Sleep and Automatically Captured Face-to-Face Interactions in Healthy Adults,” in Proceedings of the 27th Annual Meeting of the Associated Professional Sleep Societies (SLEEP), Baltimore, MD, Jun 2013.

[21] M. Butt, S. Moturu, A. Pentland, and I. Khayal, “Gender Differences in Real-Home Sleep of Young Couples,” in Proceedings of the 27th Annual Meeting of the Associated Professional Sleep Societies (SLEEP), Baltimore, MD, USA, Jun 2013.

[20] I. Khayal, M. Butt, S. Moturu, and A. Pentland, “Sociability and Sleep Quality in Healthy Adults,” in Proceedings of the 27th Annual Meeting of the Associated Professional Sleep Societies (SLEEP), Baltimore, MD, USA, Jun 2013.

[19] F. A. Baloushi, M. Butt, T. B. Ouarda, and I. Khayal, “Capturing Sleep ‘in the wild’ during the Fast¬ing month of Ramadan,” in Proceedings of the 27th Annual Meeting of the Associated Professional Sleep Societies (SLEEP), Baltimore, MD, USA, jun 2013.

[18] I. Khayal, T. B. Ouarda, M. Butt, and F. A. Baloushi, “Capturing Sleep ‘in the wild’ during the Fasting month of Ramadan,” in in Women’s Panel at Abu Dhabi Sustainability Week, Abu Dhabi, Abu Dhabi, UAE, Jan 2013.

[17] R. O. Oloritun, A. Pentland, A. Madan, and I. Khayal, “Exploring the role of duration of interac-tion in ad-hoc mobile face-to-face networks,” in Interdisciplinary Workshop on Information and Decision in Social Networks (WIDS), MIT, Cambridge, MA, USA, nov 2012.

[16] R. O. Oloritun, A. Pentland, A. Madan, and I. Khayal, “Exploring the effect of the duration and period of Face-to-Face interactions on Close Friendships,” in Interdisciplinary Workshop on Infor¬mation and Decision in Social Networks (WIDS), MIT, Cambridge, MA, USA, nov 2012.

[15] R. Oloritun, T. B. Ouarda, S. Moturu, A. Madan, A. Pentland, and I. Khayal, “Change in BMI Accu-rately Predicted via Social Exposure,” in Interdisciplinary Workshop on Information and Decision in Social Networks (WIDS), MIT, Cambridge, MA, USA, nov 2012.

[14] M. Butt, S. T. Moturu, N. Aharony, A. S. Pentland, and I. Khayal, “Effects of Alcohol on Sleep in a Real-home Environment,” in Proceedings of the 26th Annual Meeting of the Associated Professional Sleep Societies (SLEEP), Boston, MA, USA, jun 2012.

[13] S. Moturu, I. Khayal, N. Aharony, and A. S. Pentland, “Mood, Sleep and Face-to-Face Interactions in a Co-located Family Community,” in Interdisciplinary Workshop on Information and Decision in Social Networks (WIDS), Boston, MA, USA, may 2011.

12

 

REVIEWED ABSTRACTS

[12] R. Parvataneni, A. Kadambi, I. Park, A. Elkhaled, E. Essock-Burns, I. Khayal, N. Butowski, K. Lam-born, S. M. Chang, and S. J. Nelson, “Finding Early Prognostic Markers From Metabolic and Physi-ological Imaging Parameters For Newly Diagnosed GBM Patients Receiving Radiation, Temozolo-mide and A PKC Inhibitor,” in Proceedings of the 15th Annual Meeting of the Society of Neuro-Oncology (SNO), Montreal, Quebec, Canada, Nov 2010.

[11] I. Khayal, L. Jalbert, A. Elkhaled, S. M. Chang, S. Cha, and S. J. Nelson, “Evaluation of diffusion parameters as Early biomarkers of progression in high-grade gliomas,” in Proceedings of the In-ternational Society for Magnetic Resonance in Medicine (ISMRM) Seventeenth Scienti~ic Meeting, Honolulu, Hawaii, USA, Apr 2009.

[10] I. Khayal, T. McKnight, W. Bien, S. Cha, S. Chang, and S. Nelson, “ADC histogram derived RGB color-maps for characterizing low grade glioma subtypes,” in Proceedings of the International Society for Magnetic Resonance in Medicine (ISMRM) Sixteenth Scienti~ic Meeting, Toronto, Canada, May 2008.

[9] I. Khayal and S. Nelson, “Multi-exponential Diffusion Imaging: Clinically feasible Multiple B-value Diffusion Imaging for high grade gliomas,” in Proceedings of the International Society for Magnetic Resonance in Medicine (ISMRM) Sixteenth Scienti~ic Meeting, Toronto, Canada, May 2008.

[8] W. Bian, I. Khayal, C. McGue, S. Saraswathy, S. Cha, S. Chang, and S. Nelson, “Differentiation of Low-Grade Glioma Subtypes Using Mutiparametric MR Data,” in Proceedings of the International Society for Magnetic Resonance in Medicine (ISMRM) Sixteenth Scienti~ic Meeting, Toronto, Canada, May 2008.

[7] I. Khayal, T. McKnight, C. McGue, K. Lamborn, S. Chang, S. Cha, and S. Nelson, “Diffusion Imaging Parameter values aid in subtyping newly diagnosed low grade gliomas,” in Proceedings of the International Society for Magnetic Resonance in Medicine (ISMRM) Fifteenth Scienti~ic Meeting, Berlin, Germany, May 2007.

[6] I. Khayal, T. McKnight, S. Chang, S. Cha, and S. Nelson, “Apparent diffusion coefficient and frac-tional anisotropy may predict newly diagnosed low-grade glioma subtypes,” in Proceedings of the American Society of Neuro-Radiology 45th Annual Meeting. Chicago, IL: American Society of Neuro-Radiology, Jun 2007.

[5] B. Lee, I. Khayal, S. J. Chang, N. Butowski, S. Chang, S. Nelson, and S. Cha, “The utility of apparent diffusion coefficients in predicting therapeutic outcome in patients with recurrent glioblastoma multiforme treated with convection enhanced delivery of il 13-pe39qqr,” in Proceedings of the American Society of Neuro-Radiology 45th Annual Meeting. Chicago, IL: American Society of Neuro-Radiology, Jun 2007.

[4] I. Khayal, S. Cha, S. Chang, and S. Nelson, “Double b-value difference map better distinguishes tumor from edema than adc and fa maps for patients with high grade gliomas,” in Proceedings of the 2007 Joint Molecular Imaging Conference, Providence, RI, Sep 2007.

[3] I. Khayal, K. Lamborn, F. Crawford, S. Saraswathy, S. Chang, S. Cha, T. McKnight, and S. Nel¬son, “Correlation of Apparent Diffusion Coefficient and Choline in newly diagnosed patients with Glioblastoma Multiforme,” in Proceedings of the International Society for Magnetic Resonance in Medicine (ISMRM) Fourteenth Scienti~ic Meeting, Seattle, WA, USA, May 2006.

 

13

 

PRESENTATIONS (INTERNATIONAL)

[2] F. Crawford, I. Khayal, K. Lamborn, R. Choy, S. Saraswathy, I. Park, J. Oh, S. Cha, S. Chang, and S. Nelson, “Pre-surgery lactate, lipid, and adc predict survival for patients with glioblastoma mul-tiforme,” in Proceedings of the International Society for Magnetic Resonance in Medicine (ISMRM) Fourteenth Scienti~ic Meeting, Seattle, WA, May 2006.

[1] K. Karau, I. Khayal, S. Sirohey, J. Hsieh, and M. Ferrant, “Minimizing Clinical CT Variability in Sequential Volumetric Sizing of Pulmonary Nodules,” in Proceedings of the 9th International Con-ference on Screening for Lung Cancer, Jackson Memorial Hospital, Miami, FL, Oct 2003.

Op-Eds and Media

[4] M. A. Chaudhry, L. A.-R. Kadhim, I. Khayal, and A. Zaheer, “An introduction to diffusion-weighted Magnetic Resonance Imaging in oncology,” Arab Health Imaging and Diagnostics, no. 1, pp. 17–19, 2013.

[3] I. Khayal and T. B. Ouarda, “Groundbreaking Ramadan sleep study to offer fasting insight,” The National, Nov 2012. [Online]. Available: http://www.thenational.ae/news/uae-news/ groundbreaking-ramadan-sleep-study-to-offer-fasting-insight

[2] R. Ghazal, “How does Ramadan affect our sleep? UAE researchers find out,” The Na¬tional, Oct 2012. [Online]. Available: http://www.thenational.ae/news/uae-news/health/ how-does-ramadan-affect-our-sleep-uae-researchers-find-out

[1] I. Khayal, “Health is vital for the future development of the UAE,” The Na¬tional, Sep 2011. [Online]. Available: http://www.thenational.ae/news/uae-news/health/ health-is-vital-for-the-future-development-of-the-uae

Presentations (International)

[24] I. S. Khayal*, “Designing and Assessing Healthcare Systems,” Keynote Presentation at the Inter-national Conference for Axiomatic Design (ICAD2017), Iasi, Romania, September 2017.

[23] I. S. Khayal# and M. Farid, Amro, “An Architecture for a Cyber-Physical Healthcare Delivery Sys-tem with Human Agents,” The 2017 IEEE International Summer School on Smart Cities (IEEE S3C), Natal, Brazil, August, 2017.

[22] I. S. Khayal# and M. Farid, Amro, “Designing Smart Cities for Citizen Health & Well-being,” The 2017 IEEE International Summer School on Smart Cities (IEEE S3C), Natal, Brazil, August, 2017.

[21] I. Khayal# and A. M. Farid, “The Application of Model-Based Systems Engineering to the Practice of Clinical Medicine,” The 11th Annual IEEE International Systems Conference (IEEE SysCon), Montreal, Quebec, Canada, April 2017.

[20] I. Khayal#, W. Zhou, and J. Skinner, “Structuring and Visualizing Healthcare Claims Data Using Systems Architecture,” The 19th International Conference on Bioinformatics, Computational Bi-ology and Biomedical Engineering (ICBCCBE), Boston, MA, April 2017.

[19] I. Khayal# and A. Farid, “The Application of Systems Tools in the Practice of Clinical Medicine,” The 5th International Engineering Systems Symposium (CESUN2016), Washington, D.C., Jun 2016.

 

14

 

PRESENTATIONS (INTERNATIONAL)

[18] I. Khayal# and A. Farid, “Developing Sustainable Health Frameworks in Smart Cities,” 2015 IEEE Smart Cities Conference, September 2015.

[17] I. S. Khayal# and A. M. Farid, “Axiomatic Design Based Volatility Assessment of the Abu Dhabi Healthcare Labor Market: Part II - Case Study,” The Seventh International Conference on Ax-iomatic Design (ICAD2013), Worcester, MA, USA, Jun 2013.

[16] A. M. Farid and I. S. Khayal#, “Axiomatic Design Based Volatility Assessment of the Abu Dhabi Healthcare Labor Market: Part I - Theory,” The Seventh International Conference on Axiomatic Design (ICAD2013), Worcester, MA, USA, Jun 2013.

[15] M. Butt, S. T. Moturu, A. S. Pentland, and I. Khayal#, “Automatically captured sociability and sleep quality in healthy adults,” 35th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC’13), Osaka, Japan, Jul 2013.

[14] R. Oloritun, A. Pentland, and I. Khayal#, “Dynamics of Human Social Networks: People, Time, Relationships, and Places,” Collaborative Innovation Networks Conference (COINs13), Chile, Aug 2013.

[13] R. Oloritun, A. Madan, A. Pentland, and I. Khayal#, “Evolution of Social Encounters in Ad-Hoc Mobile Face-to-Face Interaction Networks,” 2012 ASE International Conference on Social Infor-matics (SocialInformatics), Alexandria, VA, USA, Dec 2012.

[12] R. O. Oloritun, A. Pentland, A. Madan, and I. Khayal#, “Identifying Close friends interactions in Face-to-Face Networks,” Applications of Social Network Analysis (ASNA2012), Zurich, Switzer-land, Sep 2012.

[11] N. Aharony, W. Pan, C. Ip, I. Khayal#, and A. Pentland, “The Social fMRI: Measuring, Understand-ing, and Designing Social Mechanisms in the Real World,” 13th International Conference on Ubiq¬uitous Computing (UbiComp 2011), Beijing, China, Sep 2011.

[10] S. T. Moturu, I. Khayal#, N. Aharony, W. Pan, and A. S. Pentland, “Using Social Sensing to Under-stand the Links between Sleep, Mood, and Sociability,” 2011 IEEE Third Intl Conference on Pri-vacy Security Risk and Trust and 2011 IEEE Third Intl Conference on Social Computing, Boston, MA, USA, Sep 2011.

[9] S. T. Moturu, I. Khayal#, N. Aharony, W. Pan, and A. S. Pentland, “Sleep, mood and sociability in a healthy population,” 2011 Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBS’11), Boston, MA, USA, Sep 2011.

[8] I. Khayal#, L. Jalbert, A. Elkhaled, S. M. Chang, S. Cha, and S. J. Nelson, “Evaluation of diffusion parameters as Early biomarkers of progression in high-grade gliomas,” Proceedings of the Inter¬national Society for Magnetic Resonance in Medicine (ISMRM) Seventeenth Scientific Meeting, Honolulu, Hawaii, USA, Apr 2009.

[7] I. Khayal#, T. McKnight, W. Bien, S. Cha, S. Chang, and S. Nelson, “ADC histogram derived RGB color-maps for characterizing low grade glioma subtypes,” Proceedings of the International Soci¬ety for Magnetic Resonance in Medicine (ISMRM) Sixteenth Scientific Meeting, Toronto, Canada, May 2008.

[6] I. Khayal# and S. Nelson, “Multi-exponential Diffusion Imaging: Clinically feasible Multiple B-value Diffusion Imaging for high grade gliomas,” Proceedings of the International Society for Mag¬netic Resonance in Medicine (ISMRM) Sixteenth Scientific Meeting, Toronto, Canada, May 2008.

15

 

PRESENTATIONS (NATIONAL)

[5] W. Bian, I. Khayal#, C. McGue, S. Saraswathy, S. Cha, S. Chang, and S. Nelson, “Differentiation of Low-Grade Glioma Subtypes Using Multiparametric MR Data,” Proceedings of the International Society for Magnetic Resonance in Medicine (ISMRM) Sixteenth Scientific Meeting, Toronto, Canada, May 2008.

[4] I. Khayal#, T. McKnight, C. McGue, K. Lamborn, S. Chang, S. Cha, and S. Nelson, “Diffusion Imag-ing Parameter values aid in subtyping newly diagnosed low grade gliomas,” Proceedings of the International Society for Magnetic Resonance in Medicine (ISMRM) Fifteenth Scientific Meeting, Berlin, Germany, May 2007.

[3] I. Khayal#, K. Lamborn, F. Crawford, S. Saraswathy, S. Chang, S. Cha, T. McKnight, and S. Nel¬son, “Correlation of Apparent Diffusion Coefficient and Choline in newly diagnosed patients with Glioblastoma Multiforme,” Proceedings of the International Society for Magnetic Resonance in Medicine (ISMRM) Fourteenth Scientific Meeting, Seattle, WA, USA, May 2006.

[2] F. Crawford, I. Khayal#, K. Lamborn, R. Choy, S. Saraswathy, I. Park, J. Oh, S. Cha, S. Chang, and S. Nelson, “Pre-surgery lactate, lipid, and adc predict survival for patients with glioblastoma mul-tiforme,” Proceedings of the International Society for Magnetic Resonance in Medicine (ISMRM) Fourteenth Scientific Meeting, Seattle, WA, May 2006.

[1] K. Karau, I. Khayal#, S. Sirohey, J. Hsieh, and M. Ferrant, “Minimizing Clinical CT Variability in Sequential Volumetric Sizing of Pulmonary Nodules,” Proceedings of the 9th International Con-ference on Screening for Lung Cancer, Jackson Memorial Hospital, Miami, FL, Oct 2003.

Presentations (National)

[24] I. Khayal*, “Applying Systems Science Methods in Health Research,” AcademyHealth’s Annual Research Meeting, New Orleans, LA, June 2017.

[23] I. Khayal#, M. McGovern, M. Bruce, and S. Bartels, “Developing an Integrated Behavioral Health System Model using Engineering Design,” The 2017 Institute of Industrial and Systems Engineer¬ing Annual Conference (IISE), Pittsburgh, PA, May 2017.

[22] I. Khayal#, M. Butt, S. Moturu, and A. Pentland, “Sociability and Sleep Quality in Healthy Adults,” The 27th Annual Meeting of the Associated Professional Sleep Societies (SLEEP), Baltimore, MD, USA, jun 2013.

[21] M. Butt, S. Moturu, A. Pentland, and I. Khayal#, “Deep Sleep and Automatically Captured Face-to-Face Interactions in Healthy Adults,” 27th Annual Meeting of the Associated Professional Sleep Societies (SLEEP), jun 2013.

[20] M. Butt, S. Moturu, A. Pentland, and I. Khayal#, “Gender Differences in Real-Home Sleep of Young Couples,” The 27th Annual Meeting of the Associated Professional Sleep Societies (SLEEP), Bal¬timore, MD, USA, Jun 2013.

[19] F. A. Baloushi, M. Butt, T. B. Ouarda, and I. Khayal#, “Capturing Sleep ‘in the wild’ during the Fasting month of Ramadan,” The 27th Annual Meeting of the Associated Professional Sleep Soci-eties (SLEEP), Baltimore, MD, USA, jun 2013.

[18] I. Khayal#, T. B. Ouarda, M. Butt, and F. A. Baloushi, “Capturing Sleep ‘in the wild’ during the Fasting month of Ramadan,” Women’s Panel at Abu Dhabi Sustainability Week, Abu Dhabi, Abu Dhabi, UAE, Jan 2013.

16

 

PRESENTATIONS (NATIONAL)

[17] I. Khayal*, “Towards Sustainable Health in Abu Dhabi,” Crown Prince Court Brown Bag Lunch, Abu Dhabi, UAE, October 3, 2012.

[16] R. O. Oloritun, A. Pentland, A. Madan, and I. Khayal#, “Exploring the role of duration of inter-action in ad-hoc mobile face-to-face networks,” Interdisciplinary Workshop on Information and Decision in Social Networks (WIDS), MIT, Cambridge, MA, USA, nov 2012.

[15] R. O. Oloritun, A. Pentland, A. Madan, and I. Khayal#, “Exploring the effect of the duration and period of Face-to-Face interactions on Close Friendships,” Interdisciplinary Workshop on Infor-mation and Decision in Social Networks (WIDS), MIT, Cambridge, MA, USA, nov 2012.

[14] R. Oloritun, T. B. Ouarda, S. Moturu, A. Madan, A. Pentland, and I. Khayal#, “Change in BMI Accu¬rately Predicted via Social Exposure,” Interdisciplinary Workshop on Information and Decision in Social Networks (WIDS), MIT, Cambridge, MA, USA, nov 2012.

[13] M. Butt, S. T. Moturu, N. Aharony, A. S. Pentland, and I. Khayal#, “Effects of Alcohol on Sleep in a Real-home Environment,” The 26th Annual Meeting of the Associated Professional Sleep Soci-eties (SLEEP), Boston, MA, USA, Jun 2012.

[12] S. Moturu, I. Khayal#, N. Aharony, and A. S. Pentland, “Mood, Sleep and Face-to-Face Interactions in a Co-located Family Community,” Interdisciplinary Workshop on Information and Decision in Social Networks (WIDS), Boston, MA, USA, May 2011.

[11] R. Oloritun, T. B. Ouarda, S. Moturu, A. Madan, A. Pentland, and I. Khayal#, “Sustainable Healthy Living: Social Interactions via embedded sensing techniques and lifestyles choices as early warn¬ing markers for Obesity,” in EPFL UAE-Swiss Research Day, Abu Dhabi-Dubai, UAE, Nov 2012.

[10] M. Butt, S. T. Moturu, N. Aharony, A. S. Pentland, and I. Khayal#, “Sustainable Healthy Living: Sociability and Sleep Quality in Healthy Adults,” in EPFL UAE-Swiss Research Day, Abu Dhabi-Dubai, UAE, Nov 2012.

[9] I. Khayal*, “Applications of Mobile Technologies in Sustainable Health,” 4th Annual SEHA Re-search Conference, Abu Dhabi, UAE, December 11, 2012.

[8] I. Khayal*, “Opportunities for Collaboration in Sustainable Health,” Visiting Innovation Profes-sor, Children’s National Medical Center, Sheikh Zayed Institute for Pediatric Surgical Innovation, Washington, D.C., May 26 2011.

[7] I. Khayal*, “Investment Opportunities in Sustainable Health Research,” Mubadala Healthcare, Mubadala Development Company, Abu Dhabi, UAE, November 27, 2011.

[6] I. Khayal*, “Innovation and Technology for Sustainable Health,” Science and Society Innovation, Technology and Biotechnology in the 21st Century Economic Opportunities and Ethical Chal-lenges, New York University Abu Dhabi, Abu Dhabi, UAE, November 16, 2011.

[5] R. Parvataneni, A. Kadambi, I. Park, A. Elkhaled, E. Essock-Burns, I. Khayal#, N. Butowski, K. Lam¬born, S. M. Chang, and S. J. Nelson, “Finding Early Prognostic Markers From Metabolic and Phys¬iological Imaging Parameters For Newly Diagnosed GBM Patients Receiving Radiation, Temo-zolomide and A PKC Inhibitor,” Proceedings of the 15th Annual Meeting of the Society of Neuro-Oncology (SNO), Montreal, Quebec, Canada, Nov 2010.

 

17

 

PRESENTATIONS (REGIONAL/LOCAL)

[4] J. Sairamesh, K. Argenbright, H. C. Simmons, R. Nemana, R. Rajagopal, I. Khayal#, and P. Anderson, “Early Warning Methods for Improving Patient Care: Harnessing Insight from Raw Patient Notes,” 3rd INFORMS Workshop on Data Mining and Health Informatics (DM-HI 2008), Washington, D.C., USA, Oct 2008.

[3] I. Khayal#, T. McKnight, S. Chang, S. Cha, and S. Nelson, “Apparent diffusion coefficient and fractional anisotropy may predict newly diagnosed low-grade glioma subtypes,” Proceedings of the American Society of Neuro-Radiology 45th Annual Meeting, American Society of Neuro-Radiology, Chicago, IL, Jun 2007.

[2] B. Lee, I. Khayal#, S. J. Chang, N. Butowski, S. Chang, S. Nelson, and S. Cha, “The utility of appar-ent diffusion coefficients in predicting therapeutic outcome in patients with recurrent glioblas-toma multiforme treated with convection enhanced delivery of il 13-pe39qqr,” Proceedings of the American Society of Neuro-Radiology 45th Annual Meeting, American Society of Neuro-Radiology, Chicago, IL, Jun 2007.

[1] I. Khayal#, S. Cha, S. Chang, and S. Nelson, “Double b-value difference map better distinguishes tumor from edema than adc and fa maps for patients with high grade gliomas,” Proceedings of the 2007 Joint Molecular Imaging Conference, Providence, RI, Sep 2007.

Presentations (Regional/Local)

[9] I. Khayal*, “Using engineering systems methods to understand variation of healthcare delivery systems at the end-of-life,” PO1 Research Meeting, Greenfield, NH, September 22, 2016.

[8] I. Khayal*, “Developing a Model & Assessment of Clinical Practice: An Engineering Perspective,” Dartmouth-Hitchock Nashua Medical Center, Nashua, NH, August 11, 2016.

[7] I. Khayal*, “Healthcare Systems Modeling and Analysis: The role of Human Resources in Abu Dhabi from 1967-2012,” The Dartmouth Institute Research in Progress Talk, Geisel School of Medicine@Dartmouth, Lebanon, NH, April 22, 2016.

[6] I. Khayal*, “Aging with Technology Support,” Tri-State Summit on Aging, Concord, NH, April 1, 2016.

[5] I. Khayal*, “Sensing Health ’in the Wild’,” Center for Technology and Behavioral Health Affiliates Meeting, Lebanon, NH, February 22, 2016.

[4] I. Khayal*, “Big Data Enabling Sustainable Health,” IBM, Abu Dhabi, UAE, October 2, 2012.

[3] K. Al Dalel and I. Khayal*, “Health and Social Networks in the Workplace and Who we are,” Pre-sentation to the Abu Dhabi Police HQ, Abu Dhabi, UAE, July 2012.

[2] I. Khayal*, “Sustainable Healthy Living,” Masdar Insitute Day at MIT, Cambrige, MA, March 1, 2011.

[1] I. Khayal*, “Nothing To Fear But Fear Itself,” Syrian-Lebanese Women’s Club of Greater Boston Scholarship Awards, Dedham, MA, USA, May 19, 2011.

 

18

 

WHITE PAPERS

Theses

[2] I. Khayal, “Characterization of Diffusion Weighted Magnetic Resonance Imaging for Patients with Brain Tumors,” Ph.D. Dissertation, University of California, Berkeley and San Francisco, 2009.

[1] I. Khayal, “Angular Dependence of Light Scattering by Apoptotic Cells,” B.Sc. Thesis, Boston Uni-versity, 2004.

White Papers

[5] I. Khayal and T. B. Ouarda, “Study of the health benefits of urban green spaces in Abu Dhabi and the United Arab Emirates,” to Mubadala Healthcare, Abu Dhabi, UAE, Tech. Rep., Apr 2012.

[4] T. B. Ouarda and I. Khayal, “Effects of climate on cardiovascular diseases in the UAE for current and anticipated conditions,” to Mubadala Healthcare, Abu Dhabi, UAE, Tech. Rep., Apr 2012.

[3] I. Khayal and T. B. Ouarda, “Characterization of MRI parameters as early warning markers of can-cer in the UAE,” to Mubadala Healthcare, Abu Dhabi, UAE, Tech. Rep., Apr 2012.

[2] I. Khayal and T. B. Ouarda, “Disease Prevention and Health Monitoring for Improved Diabetes Care in the UAE,” to Mubadala Healthcare, Abu Dhabi, UAE, Tech. Rep., Apr 2012.

[1] J. Yoo and I. Khayal, “Proactive Healthcare with Continuous Health Monitoring By Wearable Sen-sors,” to Mubadala Healthcare, Abu Dhabi, UAE, Tech. Rep., Apr 2012.

 

19

 

4 Teaching

4.1 Academic Supervision

4.1.1 Summary

Master’s Theses: [7, 6, 5, 4, 3, 2,1]

Other Advising: Research Projects & Internships: [1, 2, 3, 4, 5, 6, 7]

Post-Docs: [1]

Master’s Theses

[7] H.-W. Yang, “Environment and Health (co-advisor),” Master’s thesis, Masdar Institute of Science and Technology.

[6] R. Al Hashemi, “Environment and Health (co-advisor),” Master’s thesis, Masdar Institute of Science and Technology, 2015.

[5] K. Al Dalel, “Social Interactions and Health in the Workplace: Exploring Sickness Absence Behav-ior in the United Arab Emirates,” Master’s thesis, Masdar Institute of Science & Technology, 2014.

[4] M. Butt, “Understanding Social and Behavioral Factors Affecting Sleep in Healthy Adults,” Master’s thesis, Masdar Institute of Science & Technology, 2013.

[3] Y. Al-Abd, “UAE Aircraft Manufacturing Sector’s Potentials and Challenges (thesis defense mem-ber),” Master’s thesis, Masdar Institute of Science and Technology, 2012.

[2] R. Oloritun, “Exploring Health-Related Behaviors in Face-To-Face Networks,” Master’s Thesis, Masdar Institute of Science & Technology, 2012.

[1] D. Cihan, “Designing Robust Tuberculosis Treatment Networks in Rural India (co-advisor),” Mas-ter’s thesis, Masdar Institute of Science and Technology, 2011.

4.1.2 Other Advising: Research Projects & Internships

1. Dakota Thompson. Discrete-Event Simulation Visualization. Research Assistant Fall 2016.

2. Abdulfatai Popoola. Ramadan Sleep Study. Research Assistant Summer 2013.

3. Kefah Alqawasmeh. Practicum in Population Health Risk Assessment and Management. Visiting Student Summer 2013.

 

20

 

MASTER’S THESES

4. Hazza BaniMalek. Developing a graphical user interface design for Health Data. Summer 2012 Internship.

5. Hanin Almuhallabi. Web visualization tools for Sleep Study. Summer 2012 Internship.

6. Abdulla Al-Kayoumi. Mapping Location and Use of UAE Activity Centers. Summer 2011 Intern-ship.

7. Mouza Al Zaabi. Mapping and understanding Abu Dhabi Food Sources. Summer 2011 Internship. 4.1.3 Postdoctoral Advising

1. Ali Mutlu. Research Area: Statistical signal processing algorithms. Electrical and Computer En-gineering at Michigan State University. February 2013 - October 2013.

4.2 Curriculum Development

4.2.1 New Syllabus & Course Development

Spring 2017: CS 89/189 Health Informatics

Fall 2011: ESM 507 Management and Entrepreneurship for Engineers

Spr 2012: ESM 615 Sustainable Health

Fall 2012: Young Future Energy Leaders (YFEL) Entrepreneurship Course

Fall 2012: ESM 519 Innovation, Creativity and Entrepreneurship

Apr 2013: ESM 620 Analysis of Complex System Networks

4.3 Teaching Experience

4.3.1 Faculty Teaching Experience

Term Course Code Title Role Course

Type

Spring 2014 ESM 620 Analysis of Complex System Networks Co 

Lecturer Lecture

Fall 2013 ESM 507 Management and Entrepreneurship for Engineers Co 

Lecturer Lecture

Spring 2013 ESM 620 Analysis of Complex System Networks Lead Lecturer Lecture

Fall 2012 YFEL Entrepreneurship Course Lead Lecturer Seminar

Fall 2012 ESM 507 Management and Entrepreneurship for Engineers Lead Lecturer Lecture

Fall 2011 ESM 507 Management and Entrepreneurship for Engineers Lead Lecturer Lecture

21

 

5 Academic & Professional Service

5.1 Academic Service

5.1.1 Program & Departmental Service

The Dartmouth Institute. Geisel School of Medicine at Dartmouth.

Dates Role Scope

Summer 2016 - Member TDI Partnership Workgroup

Present

Engineering Systems & Management Department. Masdar Institute of Science & Technology.

Dates Role Scope

Fall 2013 - Fall 2014 Member MSc Admissions Committee Member

Fall 2012 - Fall 2014 Member Faculty Search Committee

5.1.2 Institutional Service

Dartmouth Hitchock Medical Center. Lebanon, NH

Dates Role Committee

Fall 2016 - Present Member D-H Population Health Initiative on Addiction & Mental

Health

Fall 2016 - Summer Member D-H Behavioral Health Planning Team - Screening &

2016 Treatment Work Group

Masdar Institute of Science & Technology. Abu Dhabi, UAE.

Dates Role Committee

Fall 2010 - Fall 2015 Chair & Founder Research Ethics Committee

Fall 2012 - Fall 2013 Member MI/MIT Joint Committee for Collaborative Research

Conference Committee

Fall 2012- Spring 2013 Member Center for Innovation Systems & Entrepreneurship

Working Group

Fall 2012- Spring 2013 Faculty Rep. Japan International Cooperation Center (JICE) Faculty

Representative

Fall 2011- Fall 2012 ESM Rep. PhD Admissions Committee

Fall 2011- Fall 2014 Member Women’s Advisory Committee

22

 

5. ACADEMIC & PROFESSIONAL SERVICE

5.2 Professional Service

5.2.1 Professional Committees

Dates Role Committee Name

Fall 2012 - Fall 2013 Member Health Authority of Abu Dhabi (HAAD) Committee on

Setting Health Research Priorities in Abu Dhabi

Fall 2012 - Fall 2013 Member Weqaya Advisory Task Force Member, Health Authority

of Abu Dhabi

Fall 2011 - Fall 2012 Judge Fifth MIT Arab Business Plan Competition, Round I&II

Fall 2011 Moderator Fifth MIT Arab Business Plan Competition Launch

Event

Fall 2010 - Fall 2011 Judge Fourth MIT Arab Business Plan Competition, Round I&II

5.2.2 Funding Review

2015 MIT Deshpande-Masdar Institute Grants

2011 MIT Energy Initiative Seed Fund Program

5.2.3 Journal Review

2016 - Present Journal of Substance Abuse Treatment

2014 - Present American Journal of Public Health

2013 - Present Journal of Enterprise Transformation

2012 - Present International Journal of Computer and Information Technology (IJCIT)

2011 - Present Psychiatric Services

2010 - Present Neuroradiology Journal

5.2.4 Conference Review

2016 - Present Council of Engineering Systems Universities (CESUN)

2012 - Present International Conference on Mobile Ubiquitous Computing Systems

5.2.5 Conference Organization

Role Scope

Panel Moderator 2017 Institute of Industrial and Systems Engineering Annual Meeting (IISE)

Panel Moderator 2017 Council of Engineering Systems Universities (CESUN) Annual Meeting

Prog. Committee 2015 IEEE Smart Cities Conference, Smart Health Track Track Chair

23

 

CIMCA 2008, IAWTIC 2008, and ISE 2008

 

Comparing Agile Software Processes Based on the Software

Development Project Requirements

Malik Qasaimeh, Hossein Mehrfard, Abdelwahab Hamou-Lhadj

Department of Electrical and Computer Engineering

Concordia University

1455 de Maisonneuve West

Montreal, Quebec H3G 1M8 CANADA

{m_qasaim, h_mehrfa, abdelw}@encs.concordia.ca

 

Abstract

Agile software development principles and methodologies have gained a lot of popularity in recent years. Unlike traditional software processes, agile processes advocate flexible practices that allow coping with changing requirements. Recently, several agile processes have been proposed, which differ in a variety of ways. In this paper, we analyze the characteristics of many agile processes and classify them based on key requirements for a software development project. The objective is to help project managers and software engineers select an agile process that is most suitable to the requirements of their software projects.

1. Introduction

Software development practices have evolved significantly since the term software engineering was popularized by F. L. Bauer during the NATO Software Engineering Conference in 1968. There exist today a large number of software processes that have been introduced and studied at a great extent, but up to date none has proven to be the golden standard.

A software process is defined as a set of activities, methods, practices, and transformations that are used to develop and maintain software and its associated products [3].

Agile processes have been proposed to overcome the flexibility issues of traditional processes. They have been developed by practitioners based on their experience working on several software development projects [8,10]. Although existing agile processes can differ significantly in the way they approach software development, they all share one key characteristic which consists of favoring close collaboration between software development and business teams via face-to-face communication, as opposed to putting an emphasis on written documentation [3, 8]. However, selecting the appropriate agile process is not an easy task due to many differences that are associated with each agile software process. These differences must be 

 

studied before choosing the appropriate methodology that fit the stakeholders’ needs [3, 4, 5]. In this paper, we surveyed and analyzed several agile processes, with the ultimate objective being:

n To help software engineers and project managers understand the strengths and weaknesses of existing agile processes.

n To help managerial staff understand the capabilities of agile processes to support software project requirements.

n To provide researchers in the area of software process engineering with a reference work in which key characteristics of several agile processes are discussed.

The remaining part of this paper is organized as follows. In Section 2, we describe the agile software processes considered in this study. In Section 3, we analyze these processes based on software development project requirements. We conclude the paper in Section 4.

2. Agile Processes

In this section, we describe the agile processes studied in this paper. We selected these particular processes in order to achieve good coverage of the types of activities available.

The processes selected in this study are the following:

n Extreme Programming (XP)

n Scrum

n Feature Driven Development (FDD)

n Adaptive Software Development (ASD)

n Crystal Methodologies (CM)

2.1. Extreme Programming (XP)

To many people, extreme programming (XP) has come to embody the agile methodology itself. This is, perhaps,

 

 

Authorized licensed use limited to: KING SAUD UNIVERSITY. Downloaded on October 4, 2009 at 07:29 from IEEE Xplore. Restrictions apply.

 

because it is one of the first agile processes that has been proposed. In general, XP consists of a set of individual practices that when put together yield a successful software practice [1].

 

Figure 1. Extreme programming (taken from [14])

Figure 1 depicts the activities involved in XP. Using XP, a development team starts by a planning phase, referred to as planning the game. During this phase, the technical team evaluates the feasibility of the project and the effort required as well as the schedule for the first release. The customer then defines the most valuable features, represented as stories, which need to be developed. Every story represents the smallest possible increment to which new features of the system can be added, which usually takes only a few weeks to be developed.

XP supports the idea that the design should be as simple as possible. To achieve this objective, XP puts an emphasis on using refactoring techniques such as removing duplicated code, improving as well as to improve the existing design. It should be mentioned that programmers must verify that the system is still operational after a refactoring activity takes place. The XP process requires that design, implementation, and testing of the system should be carried out by a pair of programmers sharing one computer. This allows programmers to spend more time finding solutions to challenging problem and less time doing routine debugging. Pair programming has been shown to be a useful technique for building robust software [7].

XP is a test driven development method such that, before writing code for a story, programmers must implement the automated tests that will verify the story functionality. As shown in Figure 1, programmers rely on unit tests to verify the correctness of the story. The work on a story is not considered complete until it has been shown to be defect free. Integration tests are run to verify that the overall functionality of the system is bug free. The story is not considered successfully implemented until it passes the acceptance test, which is normally written by the customer so as to verify that the system functionality satisfies the customer’s needs [1].

Using XP, the resulting implementation is owned by all team members. This collective ownership of the artifacts of the system allows programmers to make modifications to parts of the code that have been created by others. The main 

 

advantage of this practice is to speed up the development process such that when programmers detect a fault in the code he has the right to fix it. A coding standard is used to make sure that the development team use the same design and coding conventions. To keep the development team motivated, XP discourages team members from working more than 40 hours a week. In addition, overtime weeks are usually limited to no more than two weeks in a row [1].

2.2. Scrum

Scrum is an agile software development method that inherits many of the characteristics of the traditional iterative and incremental approaches (e.g., [1]). The term "scrum" is originated from the game of rugby, referring to a strategy of getting a ball back into play.

 

Figure 2. Steps of the Scrum process (taken from [14])

Figure 2 describes the activities involved in the Scrum process. The key activity consists of the concept of “sprint”, which consists of a 30-day working period with a set of specific goals. The Scrum process starts with a planning phase, during which a backlog list is developed to define the functionality of one or more releases of the system along with the risk associated with each release. The appropriate risk controls are also determined. The product backlog lists contain the total work of the project to be done. After that, a sprint planning meeting takes place. It usually starts every 15 to 30 days after the planning phase. During this meeting, customers, users, managers, and developers discuss the objectives of the next sprint release and the sprint backlog lists to be completed.

Scrum encourages team members to meet every day for 15 minutes in order to improve the overall knowledge of each member. Non-members may also attend but do not have the right to speak during the meeting, and any further discussions are avoided. These meetings are led by the Scrum master, a role given usually to the project manager. During these meetings, the Scrum master ensures that standards, conventions, and best practices are followed throughout the project. It is also the responsibility of the Scrum master to provide the necessary resources to team members so as to complete the sprint goals.

Scrum requires at least one daily integration and regression test of the code. In addition, a sprint review session of four hours maximum is organized regularly to discuss and report to the manager and the customer what has been

 

50

Authorized licensed use limited to: KING SAUD UNIVERSITY. Downloaded on October 4, 2009 at 07:29 from IEEE Xplore. Restrictions apply.

 

accomplished so far during the sprint. The sprint review session is also a way to receive feedback on regular basis from the various stakeholders involved in the project.

2.3. Crystal Methodologies

The Crystal methodologies are a set of processes that can be applied to different projects depending on the size and the complexity of a project. The framework in Figure 3 includes the factors that influence the selection of a particular methodology. The X-axis indicates staff size while the Y-axis represents the system criticality. The more critical the project, the more rigorous and formal processes are required. Crystal methods define four levels of critically:

n Life (L): A system failure is critical and may cause loss of life.

n Essential money (E): A system failure may cause loss of money.

n Discretionary money (D): A system failure may cause loss of money but can be fixed by referring to the system’s user manual.

n Comfort (C): A system failure may cause a loss of customer comfort.

Crystal methods put an emphasis on a set of policy standards that govern the way the project is managed. These standards are common among all crystal methods [1] and include incremental delivery of releases, progress tracking, direct user involvement, etc.

 

Figure 3. Crystal methods (taken from [14])

Different processes are assigned a different color that represents the heaviness of the process. Currently two crystal methodologies have been defined: Crystal clear and crystal orange.

Crystal clear is designed for small project with a maximum of six developers as shown in Figure 3. It can be used for different levels of criticality. For example, the D6 category 

 

indicates the use of the crystal clear method with a critical level of discretionary money.

The developers should be located in a shared space to improve the commutation between them. Developers can use any tool to improve the overall work, in other word, crystal clear keeps the choices open for developers to choose the appropriate tools. The documentation in crystal clear is very light. The requirements are expressed using UML use cases. The first incremental cycle must not exceed three months and a workshop meeting is usually held after each delivery [11].

Crystal orange is targeted for a project with a maximum of 40 developers. The project duration is usually between one to two years. The crystal orange method is suitable for a project of category D40 and may extend to E40 if necessary. Due to lack of rigorous verification techniques, the crystal orange method is not appropriate for life critical projects. Similar to crystal clear, developers are encouraged to work in a shared space. Crystal orange requires more documentation than crystal clear. For example, the requirements should be expressed in a natural language, as well as, the design document should be clear enough for developers and stockholders. The design documents are expressed using formal specifications such as state chart diagrams. The first incremental delivery must not exceed four months and more formal testing methods are encouraged in crystal orange [1, 11].

2.4. Feature-Driven Development (FDD)

The Feature-Driven Development (FDD) approach focuses on the software features of the system as the main driver of the development process. It differs significantly from the other agile processes by putting a strong emphasis on on planning and upfront design [9].

 

Figure 4. The simple steps of FDD (taken from [14])

As shown in Figure 4, the first step of the FDD process is to build a detailed model of the system to be developed, which captures the stakeholders' assumptions and requirements. Once the domain model is built, the team members print a list of the features of the system. Each feature should be developed in a few hours or days, but no longer than 2 weeks. Using FDD, development teams are formed specifically to design and implement a particular feature. The work is usually performed in parallel on multiple

 

51

Authorized licensed use limited to: KING SAUD UNIVERSITY. Downloaded on October 4, 2009 at 07:29 from IEEE Xplore. Restrictions apply.

 

features. The teams are disbanded once the feature is completed and verified. Each team is led by a feature owner, who is responsible of the code segment that implements the feature. This is contrasted with the XP approach where the ownership of the code belongs to the whole development team and not to a specific member [15].

The FDD process utilizes rigorous inspection guidelines in order to find defects in the system. It also enforces coding standards. It also encourages regular builds on a daily or weekly basis in order to add newly designed features to the baseline system. Due to the fact that features are developed in parallel, it is important to have a configuration management system that allows proper integration of the changes made to the system.

Unique to the FDD approach is a tracking and papering mechanism that assesses the project status based on the number of features that have been implemented as well as the overall progress of the design, coding, and testing activities. Each feature is scored using a value ranging between 0 (for a feature that has not yet been worked on) and 1 (a completed feature) and anything in between refers to a feature in progress.

2.5. Adaptive Software Development (ASD)

Highsmith defines the adaptive software development as “a complex adaptive process that involves interaction between agents (stockholders), environment (organization) and the product (software)” [2]. The ASD process involves essentially three phases that are geared towards planning, re¬assessment and review. These phases are: Speculate, Collaborate, and Learn. During the speculation phase a definition of what to achieve in each iteration is determined. The collaboration phase highlights the importance of team work, by sharing the knowledge among the software developers. The learning phase is carried out after each iteration in order to improve the developer’s expertise as well as to enhance the quality of the work [1].

The ASD life cycle focuses on results rather than the tasks performed to reach these results. The results are mapped directly to form the application features. In other words, the results are combined together to form the overall functionality of the system. Similar to other agile processes, ASD encourages interaction between the developers, managers, and customers. In addition, the ASD approach favors sharing of knowledge among the software developers since building a complex system would typically require the collaboration of multiple domain experts bringing the needed knowledge and experience.

Figure 5 depicts the main phases of the ASD process. ASD starts with a project initiation phase. During this phase the project mission statement is established, which is defined to guides the overall process. It must be clear and well organized. The project normally starts with unclear 

 

requirements but after each adaptive loop the overall mission becomes clearer. ASD is a feature-oriented approach rather than task-oriented. The main focus is always on the features of the systems rather than the tasks needed to implement these features. During the concurrent component engineering phase, the developers may work in parallel to implement one or more features at the same time. One of the most important aspects of ASD is the quality review phase where the customers, developers and managers meet to discuss and assess the overall quality of the work performed. The review phase session, known as the joint application development session (JAD), is important for demonstrating the functionality of the system developed as well as to keep the project within the boundaries of the mission statement. Finally, a quality assurance and release phase is held at the end of the project to fix any problems regarding the quality of the work performed [1].

 

Figure 5. ASD process activities (taken from [14])

ASD is a risk-driven process, that is, the risk management in ASD should start as early as possible in the cycle and should be implemented in all stages, i.e., Speculate, Collaborate, and Learn.

3. Requirements for Software Development Projects

In this section, we discuss key criteria that characterize software development projects against which the above software processes have been compared (see Table 1). The list of requirements is by no means exhaustive, but we believe that it captures the most common requirements found in the literature (e.g., [4, 6, 8, 12, 13]).

3.1. Customer Involvement

Customer involvement is a key practice in all agile processes, as shown in Table 1. From this table, we can clearly state that agile processes consider customers as an integral part of the development process. For example, XP, crystal methodologies, and Scrum require on-site visits to customer’s venues to allow end users to verify and prioritize the requirements during the requirement phase. The

 

52

Authorized licensed use limited to: KING SAUD UNIVERSITY. Downloaded on October 4, 2009 at 07:29 from IEEE Xplore. Restrictions apply.

 

involvement of customers is also reflected during acceptance testing, where most agile processes require these tests to be written and executed by customers.

3.2. Time to Market

Most agile processes favor early delivery of software. The motivation behind this is to allow customers to start using the software system and provide early feedback that would further enhance the upcoming releases. Early deliverables help also improve the customers’ understanding of the expected software features. Agile processes provide a delivery schedule time which varies between two weeks (e.g., in FDD) to almost six months such as in Crystal orange.

3.3. Responding to Change

As changes are costly to accommodate later in the project, the ability to respond rapidly to changes would therefore reduce project risks and costs.

The surveyed agile processes rely on some sort of review sessions involving customers to identify changes in the requirements. XP and crystal methodologies, for example, rely on frequent on-site customer visits and direct user intervention techniques to identify any possible changes triggered by customers. Scrum and ASD rely on daily meetings where customers are involved. FDD, on the other hand, identify changes during the specification of the features that need to be implemented.

3.4. Documentation

The documentation in agile processes is reduced to informal documentation, face to face communication, and on-site customer visits. Using informal documentation reduces the cost associated with maintaining it [4]. The agile processes studied in this paper vary in the level and the type of documentation they provide. For example, XP uses user stories to capture the software features that need to be implemented. Scrum’s main documentation consists of product and spring backlog lists. FDD and Crystal methodologies use UML diagrams such as use cases, class diagrams, and object models to document the design. Test cases have also been used by XP and Crystal methodologies as documentation artifacts.

3.5. Verification and Validation

Agile software processes vary in the amount of verification and validation required. This requirement will enable a project manager to select an agile process that would fit the verification and validation requirements of the software project at hand. For example, a critical safety system will 

 

most likely require extensive testing than other types of software.

Agile processes put a strong emphasis on unit and integration testing. For example, XP is known as a test-driven design approach where test cases for unit and integration testing are written before writing code. In addition to unit testing, regression testing is used in Crystal methodologies and Scrum.

Other quality review techniques are also used such as design and code inspection (in FDD), sprint review (in Scrum), and code quality review (in ASD).

3.6. Team Management

This requirement is important for organizing the team from many prospective such as team size, team communication and the use of standardized procedures (e.g., design conventions), etc. Team size is one of the important factors that may affect the selection of the development process. Although agile processes emphasize a face-to-face communication instead of formal documentation, the number of developers considered is a serious obstacle to the effectiveness of the communication. Except crystal orange, all other agile processes suggest at most 20 persons per team. Team communication considered as the second factor in team management. Agile processes tend to be people-oriented processes by allowing team members to take appropriate decisions when required without being restricted to any procedure or technique.

The use of code standard guidelines has been proposed in XP and Crystal methodologies to facilitate exchange of information among team members. This is also due to the fact that these processes favor collective ownership of the system artifacts. In other words, any member can modify the code or design of someone else. In such cases, standard coding guidelines facilitate the collaborative work.

3.7. Collaborative Work

Some software projects require coordination of project teams located in different regions, countries, and time zones. This is typically the case for large companies that have shifted parts of their software development activities to geographically distributed locations. A software process that is selected for this type of projects needs to allow for smooth collaboration among geographically distributed team members.

Agile processes do not discuss how collaborative work can be performed. This lack of support for collaborative work has also been reported by Turk et al. in [13].

 

53

Authorized licensed use limited to: KING SAUD UNIVERSITY. Downloaded on October 4, 2009 at 07:29 from IEEE Xplore. Restrictions apply.

 

Table1: Mapping agile process to software development project requirements

 

 

4. Conclusion and Future Directions

In this paper, we presented our analysis of five agile software processes and compared them based on criteria relate to software development projects. The objective is to help software engineers and project managers understand the key characteristics of these processes and therefore select the most suitable process with respect to the type of software projects they develop.

As future work, there is a need to include other agile processes not covered in this paper such as the Dynamic Systems Development Method (DSDM), the Lean Software Development (LSD), etc.

5. References

[1] P. Abrahamsson, O. Salo, J. Ronkainen, and J. Warsta, “Agile software development methods: Review and Analysis”. Espoo, Finland: Technical Research Centre of Finland, VTT Publications 478.

[2] J. A. Highsmith. Adaptive Software Development: A Collaborative Approach to Managing Complex Systems. Dorset House, New York 2000.

[3] G. Cugola and C. Ghezzi, "Software Processes: a Retrospective and a Path to the Future", In Proc. of the Software Process Improvement and Practice Conference, 1998, pp. 101-123.

[4] M.Coram and S.Bohner, "The Impact of Agile Methods on Software Project Management" In Proc. of the 12th International Conference and Workshops on the Engineering of Computer-Based Systems, 2005, pp. 363-370. 

 

[5] P. Abrahamsson, J. Warsta, M. T. Siponen, J. Ronkainen, "New Directions on Agile Methods: A Comparative Analysis", In Proc. of the International Conference on Software Engineering, 2003, pp.244- 254.

[6] M.Huo, J.Verner, L.Zhu, M.A.Babar, "Software Quality and Agile Methods”, In Proc, of the 28th Annual International Computer Software and Applications Conference, 2004, pp. 520-525.

[7] M. Fowler, K. Beck, J. Brant. Refactoring: Improving the Design of Existing Code. Addison-Wesley, 1999.

[8] S. Nerur, R. Mahapatra and G.Mangalaraj, “Challenges of migrating to agile methodologies”. Communications of the ACM, New York, 2005, pp. 72-78/

[9] S. R. Palmer and J. M. Felsing, A Practical Guide to Feature-Driven Development, Prentice Hall PTR, 2002.

[10] B. Boehm, R. Turner, “Observations on balancing discipline and agility”, In Proc. of the Agile Development Conference, 2003, pp. 32- 39.

[11] A. Cockburn. Crystal Clear: A Human-Powered Methodology for Small Teams. Addison-Wesley Professional, 2004.

[12] K. Molokken-Ostvold, K. M. Furulund, “The Relationship between Customer Collaboration and Software Project Overruns”, In Proc. of the AGILE 2007, 2007, pp 72¬83.

[13] D. Turk, R. France, and B. Rumpe, “Limitations of agile software processes”. In Proc. of the 3th International Conference on eXtreme Programming and Agile Processes in Software Engineering, 2002, pp. 43-46.

[14] D-B.Cao, “An Empirical Investigation of Critical Success Factors in Agile Software Development Projects", PhD thesis, Capella University, USA, 2006.

 

54

Authorized licensed use limited to: KING SAUD UNIVERSITY. Downloaded on October 4, 2009 at 07:29 from IEEE Xplore. Restrictions apply.

 

CV Michael Kuperberg

Personal data

Private address: Am Wald 40, 76149 Karlsruhe, Germany

Private phone: +49 721384-3825, mobile +49 163 806-4228

E-Mail: kuperberg@web.de

Citizenship: German

Family status: unmarried

Date of birth: June 23rd, 1981

Work experience

11.2010 – now Manager of the Informatics Innovation Center (IIC) at the Karlsruhe In-stitute of Technology (KIT), in charge for teaching and research on IBM System z in cooperation between KIT, IBM and FZI Karlsruhe. Involved in the acquisition of third-party funds over public-private partnerships.

03.2006 – 11.2010 Researcher and PhD candidate at the chair of Prof. Reussner, KIT. Multiple teaching and organizational responsibilities in addition to research activities.

08.2004 – 02.2006 Student research assistant (analyzing JTA transaction managers) in the Con-troMAS project at the chair of Prof. Lockemann, University of Karlsruhe.

09.2003 – 05.2004 Teaching assistant in mathematics and computer science at the UMass Boston.

08.2002 – 08.2003 Working student at SAP AG (Walldorf, Germany): working in a team de¬veloping GUI-supported Python tests for a search engine built by SAP AG.

07.2000 – 08.2002 Working student at the asknet AG in Karlsruhe, Germany. Tasks: security, smartcard and cryptography (PKCS) programming in Java.

Technologies and tools portfolio

07.2000 – now 10+ years Java programming experience incl. JTA/JTS, JNI, Reflection etc.

Software performance Palladio Component Model and its performance prediction tooling; JProfiler; VisualVM; Jamon; JavaSimon; perf4j; Dtrace and Apple Instruments

JVM analysis JMX; JConsole; JVMTI (JVM tooling interface); JIT compiler internals

Bytecode engineering ASM bytecode engineering library; Javassist; java.lang.instrument

Other languages/tools Author of TIMERMETER and BYCOUNTER; contributed to Faban; experi-ence with C#, C/C++, Scala, X10, MPI and Weka data mining framework

Dissertation and university education

03.2006 – 11.2010 Dr.-Ing. magna cum laude in Computer Science at the Department of Infor¬matics, Karlsruhe Institute of Technology (KIT). Dissertation title: “Quanti¬fying and Predicting the Influence of Execution Platform on Software Compo¬nent Performance”, Advisors: Prof. Ralf Reussner and Prof. Walter F. Tichy. 17 peer-reviewed publications (see full list starting on page 3).

10.2000 – 11.2005 Undergraduate and graduate studies in informatics (computer science) at the University of Karlsruhe, diploma grade “A” (“sehr gut”). Diploma thesis on using rule-based multi-agent systems for controling and monitoring produc¬tion scheduling. Minors in operations research and abstract algebra.

University education continues on the next page 

 

Karlsruhe, 4th May 2011 Page 1 of 4

 

CV Michael Kuperberg

08.2003 – 06.2004 Baden-W¨urttemb erg state scholarship for abroad studies (1 year) at the Uni-versity of Massachusetts in Boston as graduate student. First publication.

Professional functions, community service and memberships

2011 Session Chair at ICPE’11 (Intern. Conference on Performance Engineering).

2010 – now Elected officer and release manager of the SPEC RG (Standard Performance Evaluation Corporation Research Group), member of the steering committee.

March and April 2010 Lightning Talks “Exact Timing Measurements in Java” and “Performance measurements with perf4j” at the Java User Group Karlsruhe, Germany.

2009 – now Member of ACM, IEEE, the German Informatics Society (GI e.V.) and VKSI.

2008 Copy editor of the “Dependability Metrics” Springer LNCS volume 4909.

March 29th, 2008 Workshop host for FESCA’08 (Formal Foundations of Embedded Software and Component-Based Software Architectures), Budapest, Hungary.

2006 – now Reviewer for the IEEE TSE journal (Transactions on Software Engineering) as well as for QoSA, CBSE and other conferences and workshops.

10.2001 Tutor during the freshmen orientation phase of the computer science depart-ment of the University of Karlsruhe.

Awards and prizes

02.2011 Co-recipient of the “Best Lecture Award” for the lecture “Component-based software development” (as voted by students during the teaching evaluation at the KIT), taught during summer term 2010.

06.2009 Co-recipient of “Best Paper Award” for the QoSA 2009 conference paper “Improved Feedback for Architectural Performance Prediction using Software Cartography Visualizations” (published in Springer LNCS 5581, pages 52–69)

1999, 1997 and 1996 Prizes for Germany-wide and statewide mathematics competitions. Teaching at university level

Summer term ’11 Planning, development and teaching of two new courses at the Informatics Innovation Center (IIC), KIT: project group on software development for IBM System z and a lab course on performance analysis of applications on IBM z10.

Winter term ’10/’11 Lecturer on Scala and X10 programming languages as well as MPI (Message Passing Interface) in the lecture “Programming paradigms”.

Summer term ’10 Teaching assistant for the award-winning lecture “Component-based soft¬ware development” (voted by students as the best elective lecture of the department of computer science during the teaching evaluation at the KIT).

Winter term ’09/’10 Teaching assistant for the lecture “Software Architecture” (also winter terms ’06/’07, ’07/’08 and ’08/’09), authoring its “Middleware/SOA” part.

2006 – now Advisor of 5 diploma/master theses, 7 study/bachelor theses, 10 lab course participants and 12 seminar participants.

Karlsruhe, 4th May 2011 Page 2 of 4

 

CV Michael Kuperberg

School and college education

06.2000 Abitur at the “Goethe-Gymnasium” in Karlsruhe as top of the class, average grade “excellent” (1.0) with distinction, majors: mathematics and physics.

09.1999 – 06.2000 Coorganization of the transportation study group at the Goethe-Gymnasium.

1999 3rd prize in Germany-wide mathematics competition.

09.1997 School exchange with Lyc´ee Michelin, Marseille, France.

1997 and 1996 2nd and 3rd prize in statewide mathematics competition for high schools.

Miscellaneous

Languages: German (native language), English (fluent), French (business fluent), Russian (native language), Ukrainian (understanding and reading)

Driving license: available

Interests and hobbies: Architectural photography and history of architecture;

public transportation and high-speed trains;

playing piano; volleyball; geotagging

Peer-reviewed publications in journals and conference proceedings

Top 3 publications highlighted in red; publications listed in reverse chronological order:

1. Ginpex: Deriving Performance-relevant Infrastructure Properties Through Goal-oriented Experiments (Michael Hauck, Michael Kuperberg, Nikolaus Huber, and Ralf Reuss-ner), accepted for publication in QoSA 2011: Proceedings of the Seventh International ACM Sigsoft Conference on the Quality of Software Architectures, to appear 2011.

2. Metric-based Selection of Timer Methods for Accurate Measurements (Michael Ku-perb erg, Martin Krogmann, and Ralf Reussner), in ICPE’11: Proceedings of the 2nd ACM / SPEC International Conference on Performance Engineering, 2011.

3. Analysing the Fidelity of Measurements Performed With Hardware Performance Counters (Michael Kuperberg and Ralf Reussner), in ICPE’11: Proceedings of the 2nd ACM / SPEC International Conference on Performance Engineering, 2011.

4. Using Invariant Analysis For Improving Instrumentation-based Performance Evala-tion Of SPECjvm2008 Benchmarks (Michael Kuperberg, Martin Krogmann, and Ralf Reussner) in Proceedings of the International SPEC Benchmark Workshop 2010, 2010.

5. Using Genetic Search for Reverse Engineering of Parametric Behaviour Models for Performance Prediction (Klaus Krogmann, Michael Kuperberg and Ralf Reussner), in IEEE Transactions on Software Engineering Vol. 36, No. 6, 2010.

6. Automated Benchmarking of Java APIs (Michael Kuperberg and Fouad Omri), in Pro¬ceedings of Software Engineering 2010 (SE2010), Feb. 2010.

7. Modelling Layered Component Execution Environments for Performance Prediction (Michael Hauck, Michael Kuperberg, Klaus Krogmann and Ralf Reussner), in Proceedings of the 12th International Symposium on Component Based Software Engineering (CBSE 2009), LNCS 5582, Pages 191–208. Springer-Verlag Heidelberg, 2009.

8. Improved Feedback for Architectural Performance Prediction using Software Car¬tography Visualizations (Klaus Krogmann, Christian M. Schweda, Sabine Buckl, Michael Ku-perberg, Anne Martens and Florian Matthes), in Architectures for Adaptive Systems (Proceeding

Karlsruhe, 4th May 2011 Page 3 of 4

 

CV Michael Kuperberg

of QoSA 2009), LNCS 5581, pages 52–69. Best paper award. Springer-Verlag Heidelberg, 2009.

9. TimerMeter: Quantifying Accuracy of Software Timers for System Analysis (Michael Kuperberg, Martin Krogmann and Ralf Reussner), in Proceedings of the 6th International Con¬ference on Quantitative Evaluation of SysTems (QEST) 2009, IEEE Computer Society, Wash¬ington, DC, USA.

10. Using Heuristics to Automate Parameter Generation for Benchmarking of Java Methods (Michael Kuperberg, Fouad Omri and Ralf Reussner), in Proceedings of the 6th In-ternational Workshop on Formal Engineering approaches to Software Components and Archi¬tectures, York, UK, (ETAPS 2009, 12th European Joint Conferences on Theory and Practice of Software), March 2009.

11. Performance Prediction for Black-Box Components using Reengineered Parametric Behaviour Models (Michael Kuperberg, Klaus Krogmann and Ralf Reussner), in Proceedings of the 11th International Symposium on Component Based Software Engineering (CBSE 2008), LNCS 5282, pages 48–63. Springer-Verlag Heidelberg, 2008.

12. Reverse Engineering of Parametric Behavioural Service Performance Models from Black-Box Components (Klaus Krogmann, Michael Kuperberg and Ralf Reussner), in MDD, SOA und IT-Management (MSI 2008), Pages 57–71, GITO-Verlag, Berlin, 2008.

13. ByCounter: Portable Runtime Counting of Bytecode Instructions and Method In-vocations (Michael Kuperberg, Martin Krogmann and Ralf Reussner), in Proceedings of the 3rd International Workshop on Bytecode Semantics, Verification, Analysis and Transformation, ETAPS 2008, 11th European Joint Conferences on Theory and Practice of Software, Budapest, Hungary, April 2008.

14. Markov Models (Michael Kuperberg), in Dependability Metrics, LNCS 4909, Springer-Verlag Heidelberg, 2008.

15. Predicting Software Component Performance: On the Relevance of Parameters for Benchmarking Bytecode and APIs (Michael Kuperberg, Steffen Becker), in Proceedings of the 12th International Workshop on Component-Oriented Programming, Berlin, July 2007.

16. Influence of Execution Environments on the Performance of Software Components (Michael Kuperberg), in Proceedings of the International Research Training Groups Workshop 2006, GITO-Verlag, Berlin, 2006.

17. Metric Incremental Clustering of Nominal Data (Dan Simovici, Namita Singla and Michael Kuperberg), in Proceedings of International Conference on Data Mining 2004, Brighton, UK, 2004.

 

Karlsruhe, 4th May 2011 Page 4 of 4

 

Outline for publication

1

SCHEME OF AID AND ADVICE ON HOME MORTGAGE ARREARS

OUTLINE OF THE SCHEME

1. OVERVIEW

The Government has put in place an innovative new Scheme to help people who are struggling to pay their debts, and are at risk of losing their homes due to mortgage arrears.

The overall objective of the Scheme is to ensure that a person in this situation can access free, independent expert financial and legal advice and support, which will help them to identify and put in place their best option to get back on track. Priority is given to finding solutions which will allow the person to remain in their home, wherever that is a sustainable option.

Government Departments and bodies are working together to deliver the Scheme, which is coordinated by the Department of Justice and Equality and the Department of Social Protection, and implemented by the Money Advice and Budgeting Service (MABS), working with the Insolvency Service of Ireland (ISI), the Legal Aid Board, and the Citizens’ Information Board.

Help is already available under the Scheme, which issued its first vouchers for free advice on 22 July 2016. Contact the MABS Helpline at 076-1072000, or the local MABS office, for help and information. 

The Scheme is an important component of the new national Mortgage Arrears Resolution Service announced under the Programme for Government. It is expected to be formally launched by the Tánaiste and Minister for Justice and Equality, Frances Fitzgerald TD, and by the Minister for Social Protection, Leo Varadkar TD, in October 2016.

The Scheme is being put in place for a maximum duration of three years from July 2016, at an expected total cost of up to € 15 million. This includes funding for a significant information and promotion campaign, targeted to ensure that those most in need of help are aware of the supports available, which is expected to roll out in autumn 2016.

The Scheme also builds on recent legal changes and on previous Government initiatives to help people to recover from indebtedness and return to solvency, particularly in relation to home mortgage arrears.

2. HOW DOES THE SCHEME WORK?

The Scheme is designed to help borrowers who are insolvent and at risk of losing their homes, to get the help they need as quickly and effectively as possible. Aid and advice provided under the Scheme are provided free of charge to eligible borrowers.

 

Outline for publication

2

Anyone interested in seeking help under the Scheme should start by contacting MABS. They can call the MABS Helpline for help and information, at 076-1072000 (Monday-Friday, 9 a.m. to 8 p.m.), or call to their local MABS office (there are over 50 MABS offices around the country). Details are also available on www.keepingyourhome.ie .

MABS is the Government’s ‘one-stop shop’ point of information and guidance for homeowners in mortgage arrears, and will be the single universal gateway for aid and advice under the Scheme. All applications for aid and advice under the Scheme must go through MABS.

MABS will quickly assess the borrower’s financial situation. If the borrower is eligible for the Scheme (see section 5 below), MABS will direct them to the expert likely to be best placed to help them. Depending on the borrower’s situation, this expert may be a dedicated mortgage arrears adviser in MABS, or it may be a personal insolvency practitioner (‘PIP’), a solicitor, or an accountant. If MABS directs the borrower to a professional adviser, it will give the borrower a voucher for free provision of the relevant services.

Priority is given to obtaining financial analysis and advice, as this is the most likely to identify a long-term financial solution which can help the borrower to get out of debt and get back on track.

Once the financial analysis and advice is available, the borrower can also be referred to a ‘consultation solicitor’, for advice on any legal issues which arise.

The Scheme will also provide a new ‘duty solicitor’ service at all Circuit Court repossession hearings, to help unrepresented borrowers who are receiving advice under the Scheme.

In addition, the Scheme can provide legal aid for a borrower who considers that their creditors have unfairly refused a personal insolvency proposal which includes the home mortgage arrears, to seek review of that refusal by the Courts.

The Scheme provides five different fixed packages of professional financial or legal advice and help, all free of charge to the borrower. These are: the PIP Service, the Accountant Service, the Consultation Solicitor Service, the Duty Solicitor Service, and the Personal Insolvency Court Review Service. Each of these is explained in more detail below, at section 6.

MABS can refer a borrower to one or more of these services, depending on their needs. The intention is that a borrower in need can be routed as quickly as possible to the expert best placed to help them, and that expert advisers can work together to help the borrower to get the best available solution into place.

Advice under the Scheme can cover all options for resolving the borrower’s home mortgage arrears. Depending on the borrower’s individual circumstances, these options may include restructure, personal insolvency, bankruptcy, sale or surrender of the home. The decision on what option to take remains the borrower’s, but the Scheme aims to ensure that the borrower can make that decision with the benefit of independent expert advice.

For example, a borrower in mortgage arrears may turn out to be able to pay their debts, once they can access advice and help from MABS. Another borrower may be insolvent, yet able

 

Outline for publication

3

to keep their home under a personal insolvency arrangement or alternative restructuring arrangement, once they have access to expert financial advice. Or the borrower may not have the financial capacity to keep their home, or may not wish to keep their home, and may need independent advice and assistance on their best alternative options.

3. WHO ARE THE ADVISERS?

MABS has put in place a new network of in-house Dedicated Mortgage Arrears (DMA) advisers, who are located in MABS offices around the country. These DMA Advisers are able to provide a high level of independent expert advice to people in difficulty with mortgage arrears.

Under the Scheme, professional panels have also been established, consisting of qualified and regulated personal insolvency practitioners (‘PIPs’) 1, accountants 2 and solicitors, who have appropriate knowledge of mortgages, insolvency, bankruptcy and/or repossession. These panels are established and supervised respectively by the Insolvency Service of Ireland, the professional accountancy bodies and the Citizens’ Information Board, and the Legal Aid Board.

The professionals participating in the panels are private practitioners who have agreed to provide borrowers with independent advice and help under the Scheme, free of charge to the borrower, and at fixed cost to the Scheme. The borrower can choose their advising PIP, solicitor or accountant from the Scheme panels, subject to availability.

In order to provide help to the borrower as quickly and efficiently as possible, any adviser providing analysis and advice to them under the Scheme may share that information with any other expert who is advising that borrower under the Scheme. This will be done only with the borrower’s prior agreement, and only refers to confidential sharing of information within the Scheme, between any experts the borrower has chosen to advise them.

4. OBJECTIVES OF THE SCHEME

The detailed objectives of the Scheme are as follows:

• To provide access to independent expert advice (both financial and legal) to borrowers in serious mortgage arrears on their homes.

• To address fear and stress among many unrepresented borrowers, provide assistance to them, and ensure that sustainable solutions can be identified and put in place as early as possible.

1 PIPs are specialists in debt advice and personal insolvency solutions, and can also give financial advice on bankruptcy. They are supervised and regulated by the Insolvency Service of Ireland.

2 The solicitors’ panel and the personal insolvency practitioners’ panel are already operational. The accountancy panel is currently being established (August 2016) and is expected to become operational very shortly.

 

Outline for publication

4

• To help those in mortgage arrears who have already been issued with repossession proceedings against their homes, out of those proceedings and into alternative solutions, where that is possible.

• To address the high number of borrowers in serious home mortgage arrears, particularly those in arrears of more than 720 days, and to ensure that the home mortgage cases currently in arrears of less than 720 days are helped not to progress beyond that point.

• To support earlier, more effective and more widespread constructive engagement between borrowers and lenders where there are serious arrears on a home mortgage – right from the point where the borrower receives a proposal for surrender or other loss of the home, or is deemed ‘non-cooperating’ by the lender - so that the borrower may be assisted ‘upstream’ of the issue of repossession proceedings.

• To provide solicitor advice and assistance to those homeowners in respect of whom either surrender of ownership, or court-ordered repossession, are likely.

• To support the Courts in ensuring fair process for large numbers of otherwise unrepresented defendants in repossession actions, support MABS in dealing with cases requiring more complex financial advice, and support MABS (or PIPs on referral from MABS) in dealing with legal aspects of debt and repossession.

• To ensure access to legal aid for the new Court review under s. 115A Personal Insolvency Acts, where creditors reject a Personal Insolvency deal proposed by a Personal Insolvency Practitioner on behalf of an insolvent person, which includes mortgage arrears on their home.

5. WHO IS ELIGIBLE?

A person will be eligible for advice and assistance under the Scheme if they:

a) are insolvent (as defined under the Personal Insolvency Act 2012: this simply means that the person is ‘unable to pay their debts in full as they fall due’), and

b) are in mortgage arrears on the home in which they normally reside (their principal private residence), and

c) are at risk of losing their home (for example, they may have received from their mortgage lender repossession proceedings, a letter indicating that such proceedings will issue, a letter indicating that they are deemed non-cooperating, or an invitation to consider sale, surrender or other loss of all or part of the home).

A person will be eligible for the Personal Insolvency Court Review Service under the Scheme if they satisfy conditions (a) and (b) above, and they have already made a proposal for a Personal Insolvency Arrangement including their home, but that proposal has been rejected by their creditors. In this case, the Personal Insolvency Practitioner concerned must also certify to the Legal Aid Board that the borrower satisfies the conditions required under s. 115A of the Personal Insolvency Acts for seeking Court review.

A person will not be eligible for the Scheme, if the above conditions are satisfied, but the home is disproportionate to the reasonable living accommodation needs of the borrower

 

Outline for publication

5

and his or her dependents, taking account of the criteria set out at section 104 of the Personal Insolvency Act 2012.

The Scheme focuses on risks to the borrower’s home, and therefore does not cover a person whose mortgage arrears relate only to a buy-to-let property. However, a person in mortgage arrears on their home, who also owns a buy-to-let, will still be eligible, if he or she satisfies the conditions above.

The eligibility criteria set out above are designed to ensure that a person in serious mortgage arrears can be ‘triaged’ quickly. This will help MABS to easily identify the most suitable likely pathway to resolving their indebtedness, and to refer the borrower quickly to the appropriate expertise. The borrower’s financial situation will be examined in full detail when their Financial Statement (a detailed written analysis of the borrower’s financial situation) is being prepared. The initial triage may be revised by MABS, if information later comes to hand which makes this appropriate.

6. WHAT SORT OF ADVICE AND HELP IS COVERED?

The five services available under the Scheme are explained in detail below.

a. The PIP Service

If the borrower wishes to explore their options under personal insolvency, MABS can give them a voucher for free advice and help from a personal insolvency practitioner (‘PIP’), who is a member of the Scheme panel. The borrower can choose their PIP from the Scheme panel, subject to availability.

The voucher will entitle the borrower to a face-to-face consultation with a PIP, who will carry out a full assessment of the borrower’s financial situation, explain to them all the available options and the best option for them to deal with their mortgage arrears, and confirm that advice to the borrower in writing. As part of this service, the PIP will also prepare, with the borrower’s co-operation, a ‘Prescribed Financial Statement’ (PFS), which is the detailed written analysis of the borrower’s financial situation, as required under the Personal Insolvency Acts.

If the panel PIP advises that the borrower’s best option is personal insolvency, they will also help them in taking the next steps. If the PIP advises that the borrower’s best option is bankruptcy, they will also provide the borrower, under the voucher, with the certificate required by the bankruptcy court, confirming that the borrower has first been advised regarding their options under personal insolvency.

b. The Accountant Service

If the borrower’s case involves more complicated financial issues, but is not suited to personal insolvency, MABS may direct them for financial advice to an accountant who is a member of the Scheme panel. The borrower can choose their accountant from the Scheme panel, subject to availability.

To get free advice from an accountant under the Scheme, the borrower will need to first work with a MABS Adviser to complete the Standard Financial Statement (SFS) required by lenders under the MARP (Mortgage Arrears Resolution Process). The accountant will need a copy of this, in order to give them financial advice.

 

Outline for publication

6

The voucher will entitle the borrower to a face-to-face consultation with the panel accountant, who will advise them on any financial issues relating to resolution of their mortgage arrears and on possible solutions, and will confirm that advice to them in writing.

c. The Consultation Solicitor Service

After the borrower has completed their financial statement (SFS or PFS- see above) and received financial advice under the Scheme, their financial adviser (PIP, MABS adviser or accountant) may recommend that they also consult a solicitor from the Scheme panel, if the case raises legal issues. For example, the borrower may have received correspondence from their lender, seeking repossession or proposing surrender of their home.

The financial adviser can then apply to MABS for a legal advice voucher for the borrower. The borrower can choose their solicitor from the Scheme panel, subject to availability. The solicitor will need to see any relevant documents in the borrower’s case, and particularly their up to date SFS or PFS (see above), and any written financial advice they have received under the Scheme.

The legal advice voucher will entitle the borrower to a face-to-face consultation with a solicitor, who will assess and explain their legal position, and advise them on any repossession proceedings and any legal issues arising for resolving their mortgage arrears. The solicitor will also confirm that legal advice to the borrower in writing. If repossession proceedings have already been issued by the lender, the consultation solicitor may in certain circumstances be able to negotiate on the borrower’s behalf to settle the proceedings out of court.

d. The Duty Solicitor Service

If the borrower is facing Circuit Court repossession proceedings against their home due to mortgage arrears, they may be able to get some help at Court from the Duty Solicitor.

The Duty Solicitor is a solicitor from the Scheme’s solicitor panel, who will normally be on duty at a Circuit Court on the date where the Court is due to deal with repossession cases. The borrower cannot choose their Duty Solicitor, who will be the panel solicitor assigned by the Legal Aid Board to be present at the Court on that date.

Normally, the borrower should already have obtained written financial and legal advice under the Scheme, including advice from the consultation solicitor on any repossession proceedings. This will allow the Duty Solicitor to help the borrower more effectively.

If the borrower has not applied in time to do this, the Duty Solicitor may still be able to give them some more limited help, but only if they have already applied to MABS. Any borrower who has already received repossession proceedings against their home from their mortgage lender, and hasn’t yet applied for the Scheme, should contact MABS as soon  as possible (076 107 2000).

The Duty Solicitor Service is a limited service to provide advice and help to unrepresented borrowers at court. The borrower’s consultation solicitor will explain this in more detail.

 

Outline for publication

7

The Duty Solicitor may be able to speak for the borrower in Court, to explain what steps they are taking to try and deal with their mortgage arrears, and may be able to apply for the court proceedings to be adjourned, if the borrower is trying to put a solution in place.

The Duty Solicitor may also be able to explain to the borrower what is happening in the proceedings. However, he or she may be Duty Solicitor for a number of Scheme borrowers whose repossession cases are listed on that day, so the borrower should make sure to attend Court well in advance of the time shown on the Court papers, and should ask for MABS when they arrive. (MABS Advisers also attend at the Circuit Court on repossession hearing dates. They can provide borrowers with information and support, but they cannot provide legal advice or speak for the borrower in Court.)

The Duty Solicitor cannot act as the borrower’s legal aid solicitor, or defend the repossession proceedings on their behalf. This Scheme does not cover legal aid for defending repossession proceedings. If a borrower is in mortgage arrears on their home, the lender will normally have a legal right to seek to take back the home, and the borrower may not have a valid legal defence to prevent repossession. The consultation solicitor will advise them about this. In a small number of situations, the borrower may have a valid legal defence to repossession, but this applies to relatively few cases. If they do have a valid legal defence and want to apply for legal aid, they will need to apply for that to the Legal Aid Board, separately from this Scheme, and different conditions may apply. The consultation solicitor can explain this to the borrower.

e. The Personal Insolvency Court Review Service

The borrower will only need this service if they have already worked with a PIP, have proposed a Personal Insolvency Arrangement (‘PIA’) to their creditors, and the creditors have refused that proposal although the borrower’s PIP considers it fair and reasonable to all concerned.

Under the change made to the Personal Insolvency Acts in 2015, a borrower in this situation can ask the Courts to review the PIA proposal. If the Court agrees that the proposal is overall fair and reasonable, using the criteria set out in section 115A of the Personal Insolvency Acts, it has power to impose the proposal on the creditors.

Under the Personal Insolvency Court Review Service, the Legal Aid Board can provide the borrower with legal representation by a solicitor and barrister to make the Court review application, as part of their Civil Legal Aid service.

For this service, the borrower’s PIP applies on their behalf to the Legal Aid Board, using a Scheme reference number for the borrower provided by MABS. The PIP must also certify to the Legal Aid Board ‘that the borrower has reasonable grounds for seeking the court review and satisfies the other conditions for review laid down by the Personal Insolvency Acts.’

If the Legal Aid Board is satisfied with the application, it can then issue a Legal Aid Certificate for the borrower. The Legal Aid Board is aware of the time limits under the Personal Insolvency Acts for seeking the Court review, and will decide quickly on any fully completed application for legal aid.

 

Outline for publication

8

If the Legal Aid Board grants a Certificate, the borrower and their PIP can select a solicitor from the Scheme solicitors’ panel to act on their behalf. (Under the Personal Insolvency Acts, it is the borrower’s PIP who must apply to Court on their behalf for the review.) The Legal Aid Board will normally also cover the cost of a barrister from the Civil Legal Aid barristers’ panel. Fixed payment rates apply.

 

August 2016

 

Sri Vasavi College, Erode Self-Finance Wing 3rd February 2017

National Conference on Computer and Communication NCCC’17

http://www.srivasavi.ac.in/ nccc2017@gmail.com

Smokestack Fabric Defects Detection

Ms.P.Banumathi

Asst Professor,Department Of Computer Applications,

Tiruppur Kumaran college for Women,Tirupur

banuctech@gmail.com , 8248792149.

Dr.P.R.Tamil Selvi

Asst Professor,Department Of Computer science,

Govt Arts and Science college,Komarapalayam.

selvipr2003@gmail.com , 9942055733.

 

ABSTRACT- Quality inspection is an important aspect of modern industrial manufacturing. In textile industry Production, automate fabric inspection is important for maintain the fabric quality. For a long time the fabric defects inspection process is still carried out with human visual inspection, and thus, insufficient and costly. Therefore, Smokestack fabric defect inspection is required to reduce the cost and time waste caused by defects. The development of fully automated web inspection system requires robust and efficient fabric defect detection algorithms. The detection of local fabric defects is one of the most intriguing problems in computer vision. Texture analysis plays an important role in the automated visual inspection of texture images to detect their defects. Various approaches for fabric 

 

defect detection have been proposed in past and the purpose of this paper is to categorize and describe these algorithms.

Index Terms: Fabric Defect, Defect Classification,MATLAB,smokestack

I. Introduction

The textile industry, as with any industry today, is very concerned with quality. It is desirable to produce the highest quality goods in the shortest amount of time possible. Fabric faults or defects

are responsible for nearly 85% of the defects found by the garment industry[6]. Manufacturers recover only 45 to 65 % of their profits from seconds or off-quality goods. In this paper a fabric faulty part is taken for analysis from textiles. It is imperative, therefore, to detect, to identify, and to prevent these defects from reoccurring. There is

 

All Rights Reserved ©2017 International Journal of Computer Science (IJCS Journal) &

Department of Computer Science, Sri Vasavi College, Erode, Self-Finance Wing, Erode, Tamil Nadu, INDIA

Published by SK Research Group of Companies (SKRGC) - Scholarly Peer Reviewed Research Journals

http://www.skrgcpublication.org/ Page 1122

 

Sri Vasavi College, Erode Self-Finance Wing 3rd February 2017

National Conference on Computer and Communication NCCC’17

http://www.srivasavi.ac.in/ nccc2017@gmail.com

 

a growing realization and need for an automated woven fabric inspection system in the textile industry. All faults present on fabrics such as hole, scratch, dirt spot, fly, crack point, color bleeding etc. In this paper we analyze the faults using image processing technique. Hence the efficiency is also reduced in this process. Image processing techniques will help to production increase in fabric industry; it will also increase the quality of product. They have to detect small detail that can be located in wide area that is moving through their visual field. For this process we have use MATLAB 7.10 in imageprocessing toolbox. The high cost, along with other disadvantages of human visual inspectionhas led to the development of on-line machine vision systems that are capable of performing inspection tasks automaticallyFabric defectsIt is imperative to detect, to identify and to prevent these defects from reoccurring. There are many kinds of fabric defects[11]. Much of them are caused by machine malfunctions and havethe orientation along pick direction (broken pick yarns or missing pick yarns), they tend to 

 

be long and narrow. Other defects are caused by faulty yarns or machine spoils. Recently, the fault detection is done manually after a sufficient amount of fabric has been produced, removed from the production machine and then batched into larger rolls and then sent to the inspection frame.An optimal solution for this would be to automatically inspect from the fabric as it is being produced and to alert the maintenance personnel when the machine needs attention to preventproduction of defects or to change process parameters [12].This is done by identifying the faults in fabric using the image processing techniques and then based on the dimension of the faults; the fabric is classified and accordingly then graded. Some of the commonly occurring

fabric defects

are:

 

Fig.1 Yarn Defects & Weaving Defects

 

All Rights Reserved ©2017 International Journal of Computer Science (IJCS Journal) &

Department of Computer Science, Sri Vasavi College, Erode, Self-Finance Wing, Erode, Tamil Nadu, INDIA

Published by SK Research Group of Companies (SKRGC) - Scholarly Peer Reviewed Research Journals

http://www.skrgcpublication.org/ Page 1123

 

 

Sri Vasavi College, Erode Self-Finance Wing 3rd February 2017

National Conference on Computer and Communication NCCC’17

http://www.srivasavi.ac.in/ nccc2017@gmail.com

 

Whenever an image is converted from one form to another many types of noise can be present in the image. Here we use the Adaptive filtering to reduce stationary noise. It filters an intensity image that has been degraded by constant power additive noise. It uses pixel wise adaptive wiener method based on statistics estimated from a local neighborhood of each pixel. Thresholding

Thresholding is a process of converting a gray scale input image to a bi-level image by using an optimal threshold. The purpose of thresholding is to extract those pixels from some image which represent an object (such as graphs, maps). Though the information is binary the pixels represent a range of intensities. Here adaptive thresholding is used. In adaptive thresholding, different threshold values for different local areas are used to represents the objects.

Histogram Equalization

Histogram is a representation of the distribution of coloring an image and it represents the number of pixels that have colors in each of a fixed list of color ranges. Histogram equalization 

 

is a method for stretching the contrast by uniformly distribution the gray values enhances the quality of an image useful when the mage is intended for viewing.

V. Results

Following test image of a detected image has been used for defect identification. The image has

been exposed to histogram equalization algorithm for thresholding. The thresholding image is brought under noise removal program, where the uneven weaving is detected as spots shown in fig.3.

VI. Conclusion

The Fabric Defect detection and location identification in the normal fabrics defines the faults by this method[11]. This method classifies 85% of defect in fabric and locates the defect in the normal fabric at an acceptable rate and provides 80% classification accuracy. In the binary output image local defects appear segmented from the background. One of the most important advantages of the method is that it is multipurpose without requiring any adjustment. The versatility of the method has

 

All Rights Reserved ©2017 International Journal of Computer Science (IJCS Journal) &

Department of Computer Science, Sri Vasavi College, Erode, Self-Finance Wing, Erode, Tamil Nadu, INDIA

Published by SK Research Group of Companies (SKRGC) - Scholarly Peer Reviewed Research Journals

http://www.skrgcpublication.org/ Page 1125

 

Sri Vasavi College, Erode Self-Finance Wing 3rd February 2017

National Conference on Computer and Communication NCCC’17

http://www.srivasavi.ac.in/ nccc2017@gmail.com

 

been demonstrated not only by its applicability to different regular textures but also, for a given texture, the method allows to detect a variety of defects.

References

[1] Kaur Navneet and Dalal Mandeep “Application of Machine Vision Techniques in Textile (Fabric) Quality Analysis”, IOSR Journal of Engineering, Vol. 2, No.4, pp. 582-584, Apr. 2012.

[2] Xie Xianghua, “A Review of Recent Advances in Surface Defect Detection using Texture analysis Techniques”, Electronic Letters on Computer Vision and Image Analysis, Vol.7, No.3, pp.1-22, 2008.

[3] J.Wang, R.A. Campbell and R.J. Harwood, “Automated inspection of carpets”, in Proc. SP IE, Vol. 2345, pp. 180-191, 1995.

[4] P.M. Mahajan, S.R. Kolhe and P.M. Patil “A Review of Automatic Fabric Defect Detection Techniques” Advances in Computational Research, ISSN: 0975–3273, Vol.1, Issue 2, pp.18-29.2009.

[5] S. Priya, T. Ashok Kumar and Paul Varghese, “A Novel Approach to Fabric Defect 

 

Detection Using Digital Image Processing”, Proceedings of International Conference on SignalProcessing,Communication, Computing and Networking Technologies (ICSCCN 2011),2011.

[6] X. F. Zhang and R. R. Bresee, “Fabric defect detection and classification using image analysis”, Textile Res. J., Vol. 65, No.1, pp.1-9, 1995.

[7] E. J. Wood, “Applying Fourier and associated transforms to pattern characterization in textiles,” Textile Res. J., Vol. 60, pp. 212-220, 1990.

[8] C. Chan and G. K. H. Pang, “Fabric defect detection by Fourier analysis”, IEEE Trans. on Ind. Appl, Vol.36, No.5, pp.1267-1276, Oct 2000.

[9] T. Ashok kumar, S.O. Priya and M.G. Mini, “Optic disc localizationin ocular fundus images,” Proc. of iCVCi International Conference, India 2011.

[10] T.J. Kang et al., “Automatic Structure Analysis and Objective Evaluation of Woven Fabric Using Image Analysis”, Textile Res. J.Vol.71, No.3, pp.261-270, 2001.

 

All Rights Reserved ©2017 International Journal of Computer Science (IJCS Journal) &

Department of Computer Science, Sri Vasavi College, Erode, Self-Finance Wing, Erode, Tamil Nadu, INDIA

Published by SK Research Group of Companies (SKRGC) - Scholarly Peer Reviewed Research Journals

http://www.skrgcpublication.org/ Page 1126

 

Sri Vasavi College, Erode Self-Finance Wing 3rd February 2017

National Conference on Computer and Communication NCCC’17

http://www.srivasavi.ac.in/ nccc2017@gmail.com

 

[11] Jagrti Patel, Meghna Jain and Papiya Dutta” Detection of Faults Using Digital Image Processing Technique” 1M.Tech Scholar,

2Assistant Professor, 3Assoc. Professor,

Department of Electronics & Communication, Gyan Ganga College of Technology, Jabalpur - 482 003,Madya Pradesh, India

[12] Prof.P.Y. Kumbhar, Tejaswini Mathpati, Rohini Kamaraddiand Namrata Kshirsagar”Textile Fabric Defects Detection and Sorting Using Image Processing” international journal for 

 

research in emerging science and technology, volum-3, issue-3, mar-2016 Asian Journal of Engineering and Applied Technology ISSN 2249-068X Vol. 2 No. 1, 2013, pp.36-39© The Research Publication, www.trp.org.in

 

All Rights Reserved ©2017 International Journal of Computer Science (IJCS Journal) &

Department of Computer Science, Sri Vasavi College, Erode, Self-Finance Wing, Erode, Tamil Nadu, INDIA

Published by SK Research Group of Companies (SKRGC) - Scholarly Peer Reviewed Research Journals

http://www.skrgcpublication.org/ Page 1127

 

Gradient Pile Up Algorithm for Edge Enhancement and

Detection

Leticia Guimarães1, André Soares1,2, Viviane Cordeiro1, and Altamiro Susin1,2

1 Departamento de Engenharia Elétrica, Universidade Federal do Rio Grande do Sul - UFRGS,

Av. Osvaldo Aranha, 103, CEP. 90 035-190,Porto Alegre, RS, Brazil

leticia@eletro.ufrgs.br

http://www.lapsi.eletro.ufrgs.br

2 Instituto de Informática, Universidade Federal do Rio Grande do Sul – UFRGS,

Av. Bento Gonçalves, 9500, CEP. 91501-970, Porto Alegre, RS, Brazil

Abstract. Edge detection plays a fundamental role on image processing. The detected edges describe an object contour that greatly improves the pattern recognition process. Many edge detectors have been proposed. Most of them apply smooth filters to minimize the noise and the image derivative or gradient to enhance the edges. However, smooth filters produce ramp edges with the same gradient magnitude as those produced by noise. This work presents an algorithm that enhances the gradient correspondent to ramp edges without amplifying the noisy ones. Moreover, an efficient method for edge detection without set a threshold value is proposed. The experimental results show that the proposed algorithm enhances the gradient of ramp edges, improving the gradient magnitude without shifting the edge location. Further, we are testing the implementation of the proposed algorithm in hardware for real time vision applications.

1 Introduction

The estimation of the boundary location of an object is an important subject on image processing. Since edge detection have been studied, several techniques for edge enhancement were proposed. Most of the edge enhancement methods apply the derivative to enhance the edge of the objects on an image. Usually, the derivative of an image is produced by the convolution with a gradient operators [1], as Sobel, Roberts, Prewitt, for example.

However, there are three problems on the use of the derivative as edge detector: a) false edges produced by noise are enhanced; b) ramp edges produce lower derivative magnitude; c) the location of the ramp edges are not accurate. Several edge detection methods as proposed by Marr-Hildreth [2], Canny [3], Perona-Malik [4], have been proposed in order to provide tools to overcome the edge detection problems by applying the derivative to edge detection. Most of these methods apply a smooth gaussian operator in order to suppress the noisy edges, then, estimate the location of the edges by detection of the zero-crossing on the second derivative of an image. Edge detectors proposed by Petrou-Kitter [5] and Zang et. al. [6] are based on ramp edges model. Petrou-Kitter[5] propose a ramp edge model to an optimal edge

A. Campilho, M. Kamel (Eds.): ICIAR 2004, LNCS 3211, pp. 187–194, 2004.

© Springer-Verlag Berlin Heidelberg 2004

 

188 L. Guimarães et al.

detector. Zang et. al. [6] applies correlation matching to enhance the gradient at ramp edge location. In [7] D. Wang proposed a multiscale gradient algorithm that enhances the morphological gradient of ramp edges, however, the edge of thin objects can be lost.

Our approach aims to enhance the gradient value of blurred edges without enhances noisy gradient. In order to achieve our objective the enhancement occurs over the two component vector of the gradient, the horizontal and vertical gradient map separately. First, a smooth filter is applied to the original image. Second, the horizontal and vertical components of the gradient of an image are calculated. Third, each component of the gradient generates a connected component map, the gradient enhancement map (GEM), that guide the gradient enhancement process. Finally, the horizontal and vertical gradient map are enhanced by a process similar to a pile up, guided by their respective gradient enhancement map (GEM). The experiments presented in this paper demonstrate the improvement of the performance on edge detection using the gradient map enhanced by the proposed method.

2 Gradient Map Calculation

There are two components of the gradient, the horizontal and vertical component. The two-dimensional function gh(x,y) represents the horizontal gradient map and gv(x,y) represents the vertical gradient map of an image f(x,y), respectively. The horizontal gradient map gh(x,y) is calculated by the convolution of the image f(x,y) with a horizontal gradient operator. The vertical gradient map gv(x,y) is calculated by the convolution of the image f(x,y) with a vertical gradient operator. The gradient magnitude map g(x,y) is the magnitude of the vector composed by two components. An example of horizontal and vertical operators are the Prewitt operators.

3 Gradient Enhancement Map Generation

The gradient enhancement map GEM guides the pile up process. The pile up process occurs on two gradient maps separately following the same rules. Therefore, two gradient enhancement maps are generated, GEMH(x,y) that guides the enhancement of the horizontal gradient gh(x,y) and GEMV(x,y) that guides the enhancement of the vertical gradient gv(x,y).

The GEM is based on connected components map proposed by Bieniek and Moga

[8]. The GEM is composed by seven elements: a) zero (Z); b)right arrow ( ); c)

left arrow ( ); d) up arrow (  ); e) down arrow ; f) plateau and (P); g)

maximum (M). In [8] the connected components map is used to locate the minima, in our approach we intend to locate the maxima. Therefore, the arrows on GEM point to the maximum neighbor in magnitude. Since the direction of the arrows are related to the gradient direction, GEMH is composed only by up and down arrow and GEMV is composed by only left and right arrow.

 

4 Gradient Enhancement Process (Pile Up Process)

The gradient piled-up process is guided by the gradient enhancement maps GEM of the gradient maps, the horizontal and vertical map. The GEM indicate the direction, the start and the end points of the gradient pile up process. The arrows indicate the pile up direction. The start point of the pile up process is an arrow that the neighbor is zero (Z) and points to another arrow or maximum. The gradient value of a point at an

 

190 L. Guimarães et al.

 

(a) (b)

Fig. 1. Example of gradient enhancement process for one dimensional signal; (a) gradient enhancement map; (b) Gradient enhancement process (pile up algorithm) for the one dimensional ramp edge of the Fig. 1(a).

Search for other

START Condition

Fig. 2. Schema of the gradient enhancement process (pile up algorithm) for the vertical gradient enhancement .

 

Gradient Pile Up Algorithm for Edge Enhancement and Detection 191

 

 

(a) (b) (c)

Fig. 3. (a) Synthetic test image (60X60); (b) GEMH of the Prewitt gradient of synthetic image in (a); (c) GEMV of the Prewitt gradient of the synthetic image in (a).

arrow is piled-up over the neighbor gradient value that presents an arrow of same direction or a maximum. The process ends at a maximum. Fig. 1, shows the gradient enhancement process (pile up process) for one-dimensional ramp edge. Observe, the arrows on final GEM in Fig.1(a) are substituted by zeros (Z) as the pile up occurs, in Fig. 1(b). At the end of the pile up process, the GEM presents only maxima (M) and zeros (Z) and the gradient is enhanced from one, the original gradient value, to five, the difference between a and b level in Fig. 1(a). Moreover, the edge location is evident at the maximum on GEM. Fig. 2 shows the flowchart for the left and right pile up process to the vertical gradient enhancement.

The GEM for a synthetic image is shown in Fig. 3. Fig. 4 shows the original Prewitt gradient and the enhanced gradient for a line of the gradient map.

The edge points of edge map are those which correspond to the maximum at the GEM after the pile up process. However, only the piled up maximum is considered edge point. We intend to implement in hardware the proposed pile up algorithm, in order to accelerate the edge detection on vision chips applications.

Fig. 4. The pile up process for the line y=30 on synthetic image in Fig. 4. (a) line y=30 of the original image, Prewitt gradient and enhanced Prewitt gradient; (b) line y=30 of the noise image, Prewitt gradient and enhanced gradient.

5 Performance Evaluation

The performance of the gradient enhancement algorithm was proceeded by two methods: one applying the Pratt´s Figure Of Merit [9] FOM and other by mean absolute error MAE between the ideal edges and the detected edges.

 

192 L. Guimarães et al.

 

 

 

(a) (b) (c) (d)

Fig. 5. Test images; (a) original image; (b) image corrupted by uniform gaussian noise with variance σ=0.001; (c) edge detected at the best FOM using the Prewitt (FOM=83.66% and MAE=30.88%); (d) edge detected at the best FOM using the enhanced Prewitt gradient (FOM= 83.84% and MAE=7.99%).

The Figure of merit FOM is given by:

 

FOM 1 Na 1 (5)

 max( Na, Ni ) i1 1di2

 

where Na is the number of detected edge points, Ni is the number of ideal edge points (ground truth),  a scaling factor (we adopt  =0.1), and di distance for the ith

detected edge point.

The edge map was produced by thresholding the gradient magnitude map, however, for the original gradient the edges were thinned by the morphological skeleton operation. The FOM and MAE curve were calculated for each threshold edge map, shown in Fig. 6. The mean absolute error MAE is calculated by eq. (6).

1 M N (6)

  mi(y,x)me(y, x)

y1 x1

where mi is the ideal edge map, me is the resulting edge map, M and N are height and weight of the mi and me.

The MAE performance at the best FOM for the original Prewitt gradient and the enhance Prewitt gradient of image in Fig. 5 corrupted by gaussian noise, are shown in Fig. 7. Fig. 8 and 9 shows the results of edge detection applied to the Lena and Boat images. The ideal edge map of Lena in Fig. 8(b) was constructed manually in order to compare with the other edge detection techniques by calculating the FOM and MAE.

(a) )

(b

Fig. 6. Performance of the original and enhanced gradient for the image in Fig. 6(b); (a) Figure

Of Merit (FOM) x Threshold value ; (b) Mean Absolute Error (MAE) x Threshold value.

 

Gradient Pile Up Algorithm for Edge Enhancement and Detection 193

Fig. 7. Comparison of the Performance of Prewitt vs. Enhanced Prewitt for the image in Fig. 5.

 

 

(d) (e) (f)

Fig. 8. Edge detection on Lena; (a) original image; (b) ideal edge map of Lena; (c) best FOM

edge map from Canny with standard deviation 1 (FOM=75.28% and MAE=16.37% ); (d) smoothed Lena by mean filter 33; (e) best FOM edge map from the Prewitt gradient of Lena (FOM=79.18% and MAE=27.12%) and; (f) edge map from enhanced Prewitt (FOM=77.43% and MAE=14.72%).

 

(a) (b) (c)

Fig. 9. Edge map of the Boat image after pre-processed by a mean filter 33; (a) Original Boat image; (b) Boat image after the pre-processing; (c) the edge map generated by the proposed algorithm.

 

194 L. Guimarães et al.

6 Conclusions

The experimental results for the synthetic image show that the best FOM for the enhanced gradient is near from the best FOM of the original gradient, while for the enhanced gradient the MAE decreased 50% of the original gradient, at the best FOM threshold. Although the best FOM for original Prewitt gradient of Lena in Fig. 8, (FOM=79.18%) had been greater than the best FOM of the proposed enhanced gradient (FOM=77.43%), the MAE of the edge map obtained applying the proposed pile up method had been reduced from a MAE=27.12%, for the original gradient, to MAE=14.72% at the best FOM. Therefore, the proposed gradient enhancement method is effective to enhance gradient edges and suppress noisy gradient.

Furthermore, the FOM for the edge map produced by the proposed method is higher than the edge map generated applying the Canny edge detector, in Fig. 8(c), without setting any parameter as threshold value.

The proposed pile up process preserves the edges even for thin objects in an image, as shown in Fig. 9(c).

Acknowledgements. This work was supported by Conselho Nacional de Pesquisa CNPq – BRAZIL.

References

1. Gonzalez, R., Woods, R., Digital Image Processing, Addison-Wesley, 1992.

2. Marr, D. and Hildreth, E. C., Theory of Edge Detection, Proc. of the Royal Society of London B207, pp. 187-217, 1980.

3. Canny, J., A Computational Approach to Edge Detection, PAMI, V. 8, No. 6, pp. 679-698, 1986.

4. Perona, P. and Malik, J., Scale-Space and Edge Detection Using Anisotropic Diffusion, PAMI ,V.12 , No. 7, 1990.

5. Petrou, M. and Kitter, J., Optimal Edge Detectors for Ramp Edges, PAMI, V. 13, No. 5, pp. 483-491, 1991.

6. Wang, Z., Rao, K. R. and Ben-Arie, J., Optimal Ramp Edge Detection Using Expansion Matching, PAMI, V. 18, No. 11, pp. 1092-1097, 1996.

7. Wang, D., A Multiscale Gradient Algorithm for Image Segmentation using Watersheds, Pattern Recognition, V. 30, No. 12, pp. 2043-2052, 1997.

8. Bieniek, A and Moga, A., An efficient watershed algorithm based on connected components. Pattern Recognition, V. 33 , No. 6 , pp. 907-916, 2000.

9. Ballard, D.H., Brown, C.M., Computer Vision, Prentice Hall Inc., 1982.

 

Using Recurrent Neural Networks To Forecasting of Forex

V.V.Kondratenko1 and Yu. A Kuperin2

1 Division of Computational Physics, Department of Physics, St.Petersburg State University

2 Laboratory of Complex Systems Theory, Department of Physics, St.Petersburg State University E-mail: kuperin@JK1454.spb.edu

Abstract

This paper reports empirical evidence that a neural networks model is applicable to the statistically reliable prediction of foreign exchange rates. Time series data and technical indicators such as moving average, are fed to neural nets to capture the underlying “rules” of the movement in currency exchange rates. The trained recurrent neural networks forecast the exchange rates between American Dollar and four other major currencies, Japanese Yen, Swiss Frank, British Pound and EURO. Various statistical estimates of forecast quality have been carried out. Obtained results show, that neural networks are able to give forecast with coefficient of multiple determination not worse then 0.65. Linear and nonlinear statistical data preprocessing, such as Kolmogorov-Smirnov test and Hurst exponents for each currency were calculated and analyzed.

Key words: Neural Networks, Foreign Exchange Rate, Statistical Tests, Hurst Exponent, Complex Systems Theory

1. Introduction

Forex is the largest and most liquid of the financial markets, with an approximately $1 trillion traded every day. It leads to the serious interest to this sector of finance and makes clear that for various reasons any trader on Forex wish to have an accurate forecast of exchange rate. Most of traders use in old fashion manner such traditional method of forecast as technical analysis with the combination of fundamental one. In this paper we develop neural network approach to analysis and forecasting of financial time series based not only on neural networks technology but also on a paradigm of complex systems theory and its applicability to analysis of various financial markets (Mantegna et al., 2000; Peters, 1996) and, in particularly, to Forex. While choosing the architecture of neural network and strategy of forecasting we carried out data preprocessing on the basis of some methods of ordinary statistical analysis and complex systems theory: R/S-analysis, methods of nonlinear and chaotic dynamics (Mantegna et al., 2000; Peters, 1996). In the present paper we do not describe all of them. We present here only the results of the Kolmogorov-Smirnov test and results of R/S-analysis. However we stress that the preliminary analysis has allowed to optimize parameters of neural network, to determine horizon of predictability and to carry out comparison of forecasting quality of different currencies.

Below we give some remarks relating to advantages of neural networks technology over traditional methods and compare our approach with the methods of other authors

 

First, neural networks analysis does not presume any limitations on type of input information as technical analysis does. It could be as indicators of time series, as information about behavior of another financial instruments. It is not without foundation, that neural networks are used exactly by institutional investors (pension funds for example), that deal with big portfolios and for whom correlations between different markets are essential.

Second, in contrast to technical analysis, which is based on common recommendations, neural networks are capable to find optimal, for given financial instrument, indicators and build optimal, for given time series, forecasting strategy.

Let us remind that in present study we forecasted the exchange rates of only selected currencies on Forex market. As currencies to deal with, we chose British Pound, Swiss Frank, EURO and Japanese Yen. The following motivates this choice: practically all the main volumes of operations on Forex are made with this currencies.

Let us note that there were published a lot of papers, where similar problems have been studied. (Jingtao Yao et al., 2000; Castiglione, 2001; Kuperin et al., 2001; Lee et al., 1997; Tino et al., 2001; McCluskey, 1993). Let us briefly look round the results some of them.

In (Castiglione, 2001) there were studied the problem of sign of price increments forecasting. As an analyzed data were taken such indices as S&P500, Nasdaq100 and Dow Jones. Has been taken multilayer perceptrons of different configurations, with different number of hidden neurons. As a result there has been shown a possibility of forecasting of the sign of price increments with probability of slightly higher than 50%, i.e. a little bit better then just coin tossing. We suppose, that such kind of results is irrelevant from the practical point of view and has an academic interest.

In (McCluskey, 1993) has been studied the problem of forecasting the price increments of S&P500. As the analyzed data has been taken historical S&P500 data for the period from 1928 to 1993. There have been built and trained a lot of neural networks of different configuration. There has been estimated the profitability of using the neural networks in question. In the result surprisingly almost all networks were able to give a profitable forecast. Any statistical estimates of forecast quality were absent. It is clear that statistical estimates of forecast quality are extremely important, since profitable forecast might be just accidental.

In (Tino et al., 2001) there has been forecasted the sign of price increments. As the analyzed data there have been taken DAX and FTSE100 for some time period. The Elman recurrent neural network has been chosen. To the input there were fed binary signals corresponding to the sign of price increments. As an estimate of forecast quality, the profitability was chosen as in above paper. In the result the authors made a conclusion, that neural networks are not capable to give better results than more simple models, such as Markov models for example.

In (Jingtao Yao et al., 2000) there were carried out a neural network forecast of currencies exchange rates on Forex. As the analyzed data there were chosen weekly data of AUD, CHF, DEM, GBP and JPY. Two neural models in this study have been used:

1. Multilayer perceptron, to inputs of which there were fed values of exchange rates with some time lag.

2. Multilayer perceptron, to inputs of which there were fed values of moving average of exchange rates with different time windows (from 5 to 120 ticks).

In the result there were performed statistical estimates of forecast quality, which showed for the last model, that neural network is able to give a good result. At the same time we have to emphasize, that forecast of the exchange rate by itself is of no practical value. Besides of that, the forecast of weekly data presumes, that trader, who uses this forecast will trade once a week, which is irrelevant from the practical point of view.

We have also to mention, that forecasting of future values of underlying time series gives, besides the direct profit, the opportunity of calculating of some interesting quantities, such as price of derivative or probability of adverse mode, which is essential when one estimate the risk of investment. (Castiglione, 2001).

 

In the reminder of the paper we describe the structure of financial data we used, the data preprocessing that we carried out, i.e. normalization, calculation of Hurst exponent, Kolmogorov-Smirnov test. Next we describe the neural network architecture we used and statistical estimates of the results. Finally we present the results of time series forecasting and discussion. Last section concludes.

2. Time series forecasting

2.1 Structure of Data

In this section we remind the well-known features of economic and financial time series, which should be taken into account in training of neural networks and appropriate forecasting. What is follows are the major features:

Noise: Macroeconomic time series are intrinsically very noisy and generally have poor signal to noise ratios. The noise is due both to the many unobserved variables in the economy and to the survey techniques used to collect data for those variables that are measured. The noise distributions are typically heavy tailed and include outliers. The combination of short data series and significant noise levels makes controlling model variance, model complexity, and the bias / variance tradeoff important issues (Geman, Bienenstock and Doursat, 1992). One measure of complexity for nonlinear models is Peff , the effective number of parameters (Moody, 1992; Moody, 1994). Peff can be controlled to balance bias and variance by using regularization and model selection techniques.

Nonstationarity: Due to the evolution of the world economies over time, macroeconomic time series are intrinsically nonstationary. To confound matters, the definitions of many macroeconomic time series are changed periodically as are the techniques employed in measuring them. Moreover, estimates of key time series are periodically revised retroactively as better data are collected or definitions are changed. Not only do the underlying dynamics of the economy change with time, but the noise distributions for the measured series vary with time also. In many cases, such nonstationarity shortens the useable length of the data series, since training on older data will induce biases in predictions. The combination of noise and nonstationarity gives rise to a noise / nonstationarity tradeoff (Moody, 1994), where using a short training window results in too much model variance or estimation error due to noise in limited training data, while using a long training window results in too much model bias or approximation error due to nonstationarity.

Nonlinearity: Traditional macroeconomic time series models are linear (Granger and Newbold, 1986; Hamilton, 1994). However, recent works by some authors have suggested that nonlinearities can improve macroeconomic forecasting models in some cases (Granger and Terasvirta, 1993; Moody et al., 1993; Natter, Haefke, Soni and Otruba, 1994; Swanson and White, 1995). (See table 1 and figures 2 and 3.) Based upon our own experience, the degree of nonlinearity captured by neural network models of macroeconomic time series tends to be mild (Moody et al., 1993; Levin, Leen and Moody, 1994; Rehfuss, 1994; Utans, Moody, Rehfuss and Siegelmann, 1995; Moody, Rehfuss and Saffell, 1996; Wu and Moody, 1996). Due to the high noise levels and limited data, simpler models are favored. This makes reliable estimation of nonlinearities more difficult.

2.2 Data preprocessing

As it was mentioned above, in this paper there have been analyzed exchange rates of four main currencies on Forex: British pound, Swiss frank, EURO and Japanese Yen in the period from 22.07.1998 to 02.09.2001. Data was taken from www.gelium.net and www.forexite.com.

 

There are diagrams of exchange rates of currencies in question below for entire period and for shorter period. It could be seen that in spite of apparent smoothness graphs are rather heavily jagged.

Figure 1a. Daily exchange rate of CHF for entire period

22/07/1998 – 02/09/2001. Along the coordinate axes are given logarithmic returns

(ordinate axis) and time (abscissa axis), the definition see below

Figure 1b. Daily exchange rate of CHF for shorter period 01/01/2000– 01/09/2001. Along the coordinate axes are given logarithmic returns (ordinate axis) and time (abscissa axis), the definition see below

 

Figure 2a. Daily exchange rate of EUR for entire period 22/07/1998 – 02/09/2001. Along the coordinate axes are given logarithmic returns (ordinate axis) and time (abscissa axis), the definition see below.

Figure 2b. Daily exchange rate of EUR for shorter period 15/06/2000– 15/09/2001. Along the coordinate axes are given logarithmic returns (ordinate axis) and time (abscissa axis), the definition see below.

Figure 3a. Daily exchange rate of GBP for entire period 22/07/1998 – 02/09/2001. Along the coordinate axes are given logarithmic returns (ordinate axis) and time (abscissa axis),

the definition see below.

 

GBP

1.55

1.5

1.45

1.4

1.35

1.3

09.06.2000 09.07.2000 09.08.2000 09.09.2000 09.10.2000 09.11.2000 09.12.2000 09.01.2001 09.02.2001 09.03.2001 09.04.2001 09.05.2001 09.06.2001 09.07.2001 09.08.2001

Figure 3b. Daily exchange rate of GBP for shorter period 09/06/2000– 09/09/2001. Along the coordinate axes are given logarithmic returns (ordinate axis) and time (abscissa axis), the definition see below.

JPY

 

 

Figure 4a. Daily exchange rate of JPY for entire period 22/07/1998 – 02/09/2001. Along the coordinate axes are given logarithmic returns (ordinate axis) and time (abscissa axis),

the definition see below.

Figure 4b. Daily exchange rate of JPY for shorter period 08/10/2000–08/09/2001. Along the coordinate axes are given logarithmic returns (ordinate axis) and time (abscissa axis),

the definition see below.

Financial data has particular property: the higher is the frequency of data collection the worse is signal/noise ratio. Regard to this, weekly data are usually better forecasted, than daily

 

data. On the other hand the more accurate is the forecast the higher is its practical value. In compliance with these principles we chose daily closing prices as input data.

Usually traders are not really interested in forecast of exchange rate itself. Much more of interest is the increment of rate or the sign of increment forecast. Moreover, there is strong correlation between current and previous values of price – most probable value of price in next moment equals to its previous value. Meanwhile, in order to improve the quality of forecast it is necessary to make inputs statistically independent, i.e. to remove correlations.

Because of this in our study we analyzed not the exchange rates but its logarithmic returns:

S

r (1)

ln( )

n 

n Sn1

It is clear, that results of neural network analysis should not depend on the units of measurement. It means that all the input and output values should be reduced to a unit scale, in other words to be normalize. Moreover, in order to improve the learning rate it is effective to carry out additional data preprocessing that smoothes the data distribution before learning phase.

In the present study normalization of data was performed with the use of logistic function – the activation function of neurons. Transformation was performed corresponding to following formula:

x~

i

1  exp(

6

where: x is the mean, 6 is the standard deviation.

This transformation normalizes main volume of data and guarantees that  0,1

x~ i E .

Moreover it increases the entropy of data, since after transformation data fills the interval [0,1] more even.

The most well known but little believed hypotheses are Random Walk Hypothesis and Efficient Market Hypothesis. The Random Walk Hypothesis states that the market prices wander in a purely random and unpredictable way. The Efficient Market Hypothesis roughly speaking states that markets fully reflect all of the available information and prices are adjusted fully and immediately once new information become available. In the actual market, some people do react immediately after they have received the information while other people wait for the confirmation of information. The waiting people do not react until a trend is clearly established. All together (in week form or in stricter form) it means that in frames of Efficient Market Hypothesis is nothing to do with any type of forecasting and, in particularly, by means of neural networks. However, our study is based on the recent empirical evidence that in actual markets, the residual inefficiencies (caused by different reasons) are always present. In general we are not going to discuss these reasons here, however we shall use some of tools from linear and nonlinear statistics in order to analyze the deviation of data in question from efficiency. That is the standard Kolmogorov-Smirnov test and R/S analysis (or Hurst exponents).

The Hurst Exponent is in particular a measure of the bias in the fractional Brownian motion (Peters 1996, Peters 1994, Kantz 1997). In general (modulo any probability distribution) the method could be used in economic and financial market time series to know whether these series are biased random walk, which indicates in some sense the possibility of forecasting.

The Hurst Exponent H describes the probability that two consecutive events are likely to occur (for more details see, e.g. Peters 1996). There are three distinct classifications for the Hurst exponent: 1) H=0.5, 2) 0 <_ H < 0.5, and 3) 0.5< H <1.0. The type of series described by H=0.5 is a pure random, consisting of uncorrelated events. However, value H=0.5 cannot be used as an evidence of a Gaussian random walk. It only proves that there is no evidence of a long memory effect. A value of H different from 0.5 denotes that the observations are correlated. When 0 <_ H

 

<_ 0.5, the system is characterized by antipersistent or ergodic time series with frequent reversals and high volatility. At the same time despite the prevalence of the mean reversal concept in economic and finance literature, only few antipersistent series have yet been found. For third case (0.5 <_ H <_ 1), H describes a persistent or trend reinforcing series, which is characterized by long memory effects. The strength of the bias depends on how far H is above 0.5. However, even in the case that the Hurst process describes a biased random walk, the bias can change abruptly either in direction or magnitude. Therefore, only the average cycle length of observed data can be estimated.

It is useful before neural network forecasting to check if our time series are persistent or antipersisnet and whether returns have Gaussian distribution. This can give some information on the predictability of time series in question. If time series have H0.5 and returns have Gaussian distribution then we can presume that these time series hardly could be accurately forecasted. On the other hand, the closer H is to 1 and the more the distribution of returns differs form Gaussian distribution, the higher is the probability that these time series could be forecasted with high quality.

So to check the predictability of the analyzed time series there has been performed statistical analysis of data. For each of the currencies there were computed first four moments, i.e. mean, variance, skewness and kurtosis. Also there was performed the Kolmogorov-Smirnov test and have been computed Hurst exponents. Results are given below.

Table 1.1. Statistical data for currencies.

Mean Maximum Minimum Variance St. Deviation Skewness Kurtosis

CHF -8,3E-05 -0,03798 0,035216 4,92E-05 0,007017 0,2380862,291049

EUR 0,000246 -0,02335 0,023201 3,2E-05 0,005661 -0,18691 2,138603

GBP -1,7E-05 -0,02245 0,024826 2,42E-05 0,004917 0,009823 1,655181

JPY 4,13E-05 -0,07811 0,035251 6,01E-05 0,007755 -0,94684 7,928037

Table 1.2 The Kolmogorov-Smirnov test results.

CHF 0.05871

EUR 0.010531

GBP 0.04967

JPY 0.06328

Table 1.3 The values of Hurst Exponents

CHF 0.566  0.002

EUR 0.570  0.002

GBP 0.523  0.004

JPY 0.557  0.002

As it is shown in Tables 1.1-1.3, the values of Hurst exponents for the logarithmic returns of daily exchange rates data is higher than 0.5 for all the observed time series. This shows that all the studied currencies are persistent. The highest value is 0.570 for the exchange rate of EUR/USD. In other words, it is the least efficient currency amongst the four studied ones. The results of the Kolmogorov-Smirnov test show that not any returns of observed time series have Gaussian distribution. This means that not any time series are described by Brownian motion. The closest distribution to Gaussian is the distribution of EUR logarithmic returns. It means, that EUR returns are weaker correlated than the others currencies time series returns and probably EUR will have the worst forecast.

 

The results obtained in this part of the study imply that the all time series are nor antipersistent neither Brownian motion. Thus neural network forecasting of the exchange rates is possible and should has a good quality.

 

Figure 5. Distribution of CHF. Here abscissa axis is the upper boundaries of fragmentation of

the logarithmic returns and ordinate axis is the number of observations of return’s value in each

fragmentation interval.

 

 

Figure 6. Distribution of EUR. Here abscissa axis is the upper boundaries of fragmentation of the

logarithmic returns and ordinate axis is the number of observations of return’s value in each

fragmentation interval.

 

Figure 7.Distribution of GBP. Here abscissa axis is the upper boundaries of fragmentation of the

logarithmic returns and ordinate axis is the number of observations of return’s value in each

fragmentation interval.

 

 

Figure 8.Distribution of JPY. Here abscissa axis is the upper boundaries of fragmentation of the

logarithmic returns and ordinate axis is the number of observations of return’s value in each

fragmentation interval.

1.3 Choosing the of neural network architecture

The presence of recurrent feedback in neural network seems to be a positive factor in forecasting of financial time series. (Castiglione, 2001; McCluskey, 1993) This is apparently because the recurrent neural network has a “deeper memory” than feedforward neural network.

In our study the Elman-Jordan neural network has been used. This class of neural networks with learning by back propagation method was successfully used for prediction of financial markets since it is the recurrent neural net which learns the rules in the time series, which is necessary when works with ones. The disadvantage of this class of neural nets is the long learning time. (Kuperin et al., 2001).

The network we built had two inputs and one output. To one input there were fed daily returns, to another one there were fed daily returns, smoothed by moving average with the window equal to 5 observations. Moving average in input was supposed to smooth the noisy time series of returns. As the output we chose the value of moving average of returns with the window equal to 5 observations, shifted to one day forward, because it corresponds to trading purposes better than returns because of more accurate forecast and such fact that if trader would try to trade on every market movement it would lead to very expensive transaction costs., It is partially explain why it is reasonable to forecast the moving average. Thus, the network was supposed to forecast the one-day ahead value of moving average of returns with the window equal to 5 observations. The number of hidden neurons was chosen equal to 100. .

Neurons of input layer had linear activation functions. Neurons of hidden and output layers had logistic activation function. This configuration was determined experimentally as giving the best results.

 

Momentum and learning rate were chosen equal to 0.003. The net terminated its learning if during 20000 epochs it was not capable to improve its results. Training, test and production sets were chosen as following: first 1000 observations were used as the training set, next 200 observations were used as the test set, and last 100 observations were used as the production set.

 

Figure 9. Schematic illustration of the Elman-Jordan network architecture. Here r(t) is

logarithmic return, MA(r,5) is the moving average with window equal to 5 of logarithmic returns

and Loop Back denotes the interlayer feedback.

2.4 Statistical estimation of forecasting results

In the present section we briefly describe some indices we used for statistical estimations of the neural network forecasting quality. This part of the study is quite important both for the proper choice of neural network architecture and for statistical relevance of obtained results. The specific values of obtained indices are discussed below.

In particular, as performance criteria there were chosen the following indices:

R2 The coefficient of multiple determination is a statistical indicator usually applied to multiple regression analysis. It compares the accuracy of the model to the accuracy of a trivial benchmark model wherein the prediction is just the mean of all of the samples. A perfect fit would result in an R squared value of 1, a very good fit near 1, and a very poor fit less than 0. If the neural model predictions are worse than one could predict by just using the mean of the sample case outputs, the R squared value will be less than 0. The formula used for R2 is given by

 ( y  y ˆ)

i 2

R 21

2

i

where: yi is the actual value, yˆ is the predicted values of yi , y is the mean of the yi values

r squared This is the square of the correlation coefficient, described below.

Mean Squared Error This is the mean over all patterns of the square of the actual value

minus the predicted value, i.e., the mean of (actual minus predicted) squared.

 

Min Absolute Error This is the minimum of actual minus predicted values over all patterns.

Max Absolute Error This is the maximum of actual minus predicted values over all patterns.

Correlation Coefficient r (Pearson’s Linear Correlation Coefficient) This is a statistical measure of the strength of the correlation between the actual and predicted outputs. The r coefficient can range from -1 to +1.

Percent within 5%, 10%, 20% and 30% and over 30% It is the percent of network answers that are within the specified percentage of the actual answers used to train the network. If the actual answer is 0, the percent cannot be computed and that pattern is not included in a percentage group. For that reason and rounding, the total computed percentages may not add up to 100%.

Ability of correct prediction for the sign of increments This is the number of correctly predicted signs of increments, divided by the whole number of patterns.

Most representative statistical estimate of forecast quality among of all mentioned above is R2. Other representative indices are: r squared, mean absolute error and maximum absolute error. Mean absolute error and maximum absolute error show the accuracy of forecast, r squared shows the linear correlation, but tells nothing about accuracy. R2 gives information both on accuracy and correlation. It can be shown that

 (1 2 ) 

R (4)

where:  is the mean absolute error,  is the standard deviation.

For example, if we have R2 = 0.8, so we can estimate mean absolute error as  0.447 .

3. Results and discussion

In the present study there were built and trained four neural networks, one for each currency. In Appendix we present the Tables with the statistical results of forecast quality.

Below we give also the examples of graphs, which visualize the quality of forecast, where ordinate axis is the value of logarithmic returns and abscissa axis is the time.

Figure 10. Actual and predicted values of CHF for the period from 18/04/2001 to

01/10/2001.

 

Figure 11. Actual and predicted values of EUR for the period from 12/08/2001 to

04/05/2002.

Figure 12. Actual and predicted values of GBP for the period from 18/04/2001 to

01/10/2001.

Figure 13. Actual and predicted values of JPY for the period from 23/04/2001 to

01/10/2001.

 

We also made some attempts to improve the achieved results in the following direction. Since the amount of data in use was limited by available data on EURO (exchange rate of EURO appeared in 1998), we decided to take more data in spite of its different amount for different currencies. Thus, the amount of data in use was chosen equal to 2200 daily ticks. Moreover, we decided to change the configuration of inputs and to add some technical indicators to the input of the neural net. We have chosen the Relative Strength Index (RSI) and Williams Percent Range (%R) as such inputs. In other words there were built several neural networks with different configuration of inputs. Below we give more detailed description of these models. (See Appendix for the appropriate results and statistical estimation). Namely, we undertaken the following numerical experiment:

1. We have added to the input of neural network such technical indicators as RSI and %R. We expected to improve the quality of forecast due to presence of leading technical indicators in the network input. Since the number of inputs increased, we also increased the number of hidden neurons up to 140.

2. As the next step we have decided to feed only technical indicators in input of the network. Moving average of returns and returns themselves were removed. We supposed to check the prevalent belief that all the necessary information contains in RSI and % R time series and so the excess information could only worsen the quality of forecast.

3. Since it turned out that the results, shown by previous model were essentially worse than the others, we decided to feed to the network input the technical indicators and moving average of returns, removed at the same time returns themselves from the input. We supposed that presence of raw data in the input could worsen the quality of forecast.

4. Since the best results have been shown by the network, which had two indicators, i.e., moving average and returns in input, we have decided to take this model as the basic one and only slightly modify it. We replaced moving average by exponential moving average, since it approximates time series more accurate and contains information on longer time period.

5. Finally, we have used the embedding method. Namely, we fed four series values of returns and supposed to obtain the next day returns from the output.

4.Conclusions

Let us summarize the obtained results as follows.

The statistically best results have been obtained by the model, where logarithmic returns and moving average were taken as inputs. Comparing the results, obtained from this model, to others ones it should be noted that:

1. Model with two added technical indicators: in general, the forecast quality became worse than the initial model (R2 is increased on CHF only, on the others currencies it decreased). It turned out that the probability of correctly predicted increments sign is also lower. Apparently, the information that contains in indicators is excess and makes the problem of forecast more complicated. We could increase, of course, the number of hidden neurons, but it would lead to unwarranted increasing of the learning time.

2. Model with only RSI and %R in input: statistically the forecast quality fell down even more then in other models we studied. This indicates, that the mentioned

 

above prevalent belief about information content of technical indicators is irrelevant to the neural network forecasting.

3. Model with RSI, %R and moving average in input: comparing to previous models the forecast quality has been improved, but it still remained not the best. It looks like the presence of moving average in input of the network is the positive factor at least due to the smoothing of the data.

4. Model with indicators, returns and exponential moving average: results, obtained from this model are analogous to results obtained by basic model. Apparently, there is we have found no essential difference between ordinary moving average and exponential moving average for the problem in question.

5. Embedding method: the particular property of this model is that the results of forecasting of different currencies are much more analogous, than in other models. Nevertheless the forecast quality is worse than in basic model. Apparently, it is possible to obtain better results from this model by increasing the number of inputs and hidden neurons, but it would lead again to unjustified long learning time.

It should be noted, that CHF and JPY are usually better predicted than GBP and EURO. Is seemingly means that regularities, that contain CHF and JPY time series are more complicated than regularities, containing in GBP and EURO time series. In EURO case comparatively lower quality of forecast could be explained by insufficient amount of data.

Thus, in this study neural net forecast of four main currencies on Forex market was carried out. Finally let us conclude that statistical estimates of Forex forecast indicates, that neural network can predict the increments sign with relatively high probability – approximately 80%, which is quite sufficient for practical use. The coefficient of multiple determination R2 in the best our model varied from 0.65 to 0.69. . We believe that such quality of forecast is enough for building an automatic profitable trade strategy. This point of view has been confirmed in (Kuperin et al., 2001) where a profitable trade strategy for GBP, based on forecast with similar quality has been constructed and tested.

 

References

1. R.N.Mantegna, H.E.Stenley: An Introduction to Econophysics. Correlations and Complexity in Finance. Cambridge University Press (2000)

2. E.E.Peters: Chaos and Order in Capital Markets. John Wiley&Sons (1996)

3. E.E.Peters: Fractal Market Analysis. John Wiley&Sons (1994)

4. H.Kanz, T.Schreiber: Nonlinear Time Series Analysis. Cambridge University Press (1997)

5. A.A.Ezhov, S.A. Shumsky. Neurocomputing and its Applications in Economics and Business. Moscow.: MEPHI, 1998.

6. Jingtao Yao, Chew Lim Tan. A Case Study on Using Neural Networks to Perform Technical Forecasting of Forex. Neurocomputing, Vol. 34, No. 1-4, pp79-98 (2000).

7. Filippo Castiglione. Forecasting Price Increments Using an Artificial Neural Network. in Complex Dynamics in Economics, a Special Issue of Advances in Complex Systems, 3(1), 45-56, Hermes-Oxford (2001)

8. Kuperin, Yuri A., Dmitrieva Ludmila, A., Soroka Irina V. Neural Networks in financial markets dynamical studies: Working paper series of the Center for Management and Institutional Studies. School of Management. Saint-Petersburg State University. No 2001-12

9. C.Lee Giles, Steve Lawrence, Ah Chung Tsoi. Noisy Time Series Prediction using a Recurrent Neural Network and Grammar Inference. Proceedings of IEEEE/IAFE Conference on computational intelligence for financial engineering IEEE, Piscantaway. New York, (1997). P. 253- 259.

10. John Moody. Economic Forecasting: Challenges and Neural Network Solutions. Proceedings of the International Symposium on Artificial Neural Networks. Hsinchu, Taiwan, (1995)

11. Peter C. McCluskey. Feedforward and Recurrent Neural Networks and Genetic Programs for Stock Market and Time Series Forecasting. Master's thesis, Brown University, (1993).

12. Peter Tino, Christian Schittenkopf, Georg Dorffner. Financial Volatility Trading using Recurrent Neural Networks. Pattern Analysis and Application, (2001)

13. Moody, J. The effective number of parameters: an analysis of generalization and regularization in nonlinear learning systems, in J. E. (1992)

14. Moody, J., Challenges of Economic Forecasting: Noise, Nonstationarity, and Nonlinearity, Invited talk presented at Machines that Learn, Snowbird Utah, April (1994a).

15. Moody, J., Prediction risk and neural network architecture selection, in V. Cherkassky, J. Friedman and H. Wechsler, eds, `From Statistics to Neural Networks: Theory and Pattern Recognition Applications', Springer-Verlag. (1994b)

16. Moody, J., Levin, A. and Rehfuss, S., `Predicting the U.S. index of industrial production', Neural Network World 3(6), 791--794. Special Issue: Proceedings of Parallel Applications in Statistics and Economics '93. (1993)

17. Natter, M., Haefke, C., Soni, T. and Otruba, H., Macroeconomic forecasting using neural networks, in `Neural Networks in the Capital Markets 1994'. Pi, H. and Peterson, C. (1994), `Finding the embedding dimension and variable dependencies in time series', Neural Computation pp. 509--520. (1994)

18. Swanson, N. and White, H., A model selection approach to real-time macroeconomic forecasting using linear models and artificial neural networks, Discussion paper, Department of Economics, Pennsylvania State University. (1995)

19. Rehfuss, S., Macroeconomic forecasting with neural networks. Unpublished simulations. (1994)

20. Utans, J., Moody, J., Rehfuss, S. and Siegelmann, H., Selecting input variables via sensitivity analysis: Application to predicting the U.S. business cycle, in `Proceedings of Computational Intelligence in Financial Engineering'. (1995)

21. Wu, L. and Moody, J., `A smoothing regularizer for feedforward and recurrent networks', Neural Computation 8(2). (1996)

 

Appendix

Table 2. Statistics of forecast quality for main currencies of Forex (the basic model)

CHF EUR GBP JPY

Number of epochs 2622 27370 14389 10181

Number of patterns in training set: 1200 1200 1200 1200

Number of patterns in production set: 103 103 103 103

R squared: 0,6529 0,6694 0,6714 0.6901

r squared: 0,6550 0,6697 0,6722 0.6917

Mean squared error: 0,003 0,001 0,003 0.002

Mean absolute error: 0,041 0,029 0,041 0.039

Minimum absolute

error: 0 0 0,001 0

Maximum absolute

error: 0,178 0,122 0,146 0.150

Pearson’s correlation coefficient: 0,8093 0,8183 0,8199 0.8317

Percent within 5%: 37,879 52,427 35,960 39.798

Percent within 5% to 10%: 28,283 24,272 30,505 30.101

Percent within

10% to 20%: 26,010 23,301 26,263 24.848

Percent within 20% to 30%: 5,556 0 5,657 3.636

Percent over 30%: 2,273 0 1,616 1.616

Probability of correct

sign of increments

forecasting: 0,82 0,82 0,77 0.83

 

Table 3. Statistics of forecast quality for main currencies of Forex for the network with added technical indicators

CHF EUR GBP JPY

Number of epochs 3921 149 2404 12581

Number of patterns in training set: 1600 1200 1600 1600

Number of patterns in production set: 100 100 100 100

R squared: 0.6836 0.4530 0.5436 0.6168

r squared: 0.6941 0.4584 0.5503 0.6212

Mean squared error: 0.002 0.002 0.003 0.002

Mean absolute error: 0.037 0.038 0.040 0.034

Minimum absolute

error: 0 0.001 0.001 0

Maximum absolute

error: 0.155 0.110 0.139 0.137

Pearson’s correlation

coefficient: 0.8332 0.6770 0.7418 0.7882

Percent within 5%: 45.299 50.000 35,960 39.655

Percent within 5% to 10%: 22.222 20.192 29.310 34.483

Percent within

10% to 20%: 26.496 21.154 25.000 22.414

Percent within 20% to 30%: 3.419 7.692 6.897 2.586

Percent over 30%: 2.564 0.962 0.862 0.862

Probability of correct

sign of increments

forecasting: 0,78 0,78 0,73 0.81

 

Table 4. Statistics of forecast quality for main currencies of Forex for the network with two added indicators and without moving average and returns.

CHF EUR GBP JPY

Number of epochs 1918 114 6271 136

Number of patterns in training set: 1600 1200 1600 1600

Number of patterns in production set: 100 100 100 100

R squared: 0.5153 0.4032 0.4114 0.4693

r squared: 0.5332 0.4061 0.4139 0.4798

Mean squared error: 0.003 0.003 0.003 0.003

Mean absolute error: 0.045 0.039 0.047 0.041

Minimum absolute

error: 0 0 0 0

Maximum absolute

error: 0.168 0.130 0.142 0.142

Pearson’s correlation

coefficient: 0.7302 0.6373 0.6433 0.6927

Percent within 5%: 37.607 42.308 29.310 36.207

Percent within 5% to 10%: 23.077 24.038 33.621 31.897

Percent within

10% to 20%: 26.496 25.000 28.448 24.138

Percent within 20% to 30%: 7.692 7.692 5.172 7.759

Percent over 30%: 5.128 0.962 3.448 0

Probability of correct

sign of increments

forecasting: 0,74 0,75 0,73 0,80

 

Table 5. Statistics of forecast quality for main currencies of Forex for the network

without raw data in input

.

CHF EUR GBP JPY

Number of epochs 3794 149 1885 1997

Number of patterns in training set: 1600 1200 1600 1600

Number of patterns in production set: 100 100 100 100

R squared: 0.6513 0.4541 0.5124 0.6022

r squared: 0.6614 0.4601 0.5177 0.6051

Mean squared error: 0.002 0.002 0.003 0.002

Mean absolute error: 0.039 0.038 0.042 0.036

Minimum absolute

error: 0 0 0 0

Maximum absolute

error: 0.166 0.108 0.140 0.126

Pearson’s correlation

coefficient: 0.8132 0.6783 0.7195 0.7779

Percent within 5%: 39.316 48.077 37.931 41.379

Percent within 5% to 10%: 27.350 21.154 27.586 31.897

Percent within

10% to 20%: 25.641 23.077 26.724 24.138

Percent within 20% to 30%: 5.128 6.731 6.897 2.586

Percent over 30%: 2.564 0.962 0.862 0

Probability of correct

sign of increments

forecasting: 0,78 0,76 0,71 0,81

 

Table 6. Statistics of forecast quality for main currencies of Forex for the network with exponential moving average in input.

CHF EUR GBP JPY

Number of epochs 4143 132 3518 117

Number of patterns in training set: 1600 1200 1600 1600

Number of patterns in production set: 100 100 100 100

R squared: 0.6825 0.4470 0.5106 0.6954

r squared: 0.6841 0.4384 0.5116 0.7024

Mean squared error: 0.002 0.002 0.003 0.001

Mean absolute error: 0.037 0.038 0.042 0.032

Minimum absolute

error: 0.001 0.001 0 0

Maximum absolute

error: 0.128 0.110 0.144 0.084

Pearson’s correlation

coefficient: 0.8271 0.6580 0.7152 0.8381

Percent within 5%: 40.171 50.000 35.345 40.517

Percent within 5% to 10%: 31.624 20.192 31.897 39.655

Percent within

10% to 20%: 22.222 21.154 25.000 18.103

Percent within 20% to 30%: 3.419 7.692 6.034 1.724

Percent over 30%: 2.564 0.962 1.724 0

Probability of correct

sign of increments

forecasting: 0,79 0,78 0,78 0,89

 

Table 7. Statistics of forecast quality for main currencies of Forex for the network with four values of returns in series

CHF EUR GBP JPY

Number of epochs 2282 1764 3517 3513

Number of patterns in training set: 1600 1200 1600 1600

Number of patterns in production set: 100 100 100 100

R squared: 0.6369 0.5946 0.5103 0.5678

r squared: 0.6476 0.6030 0.5212 0.5706

Mean squared error: 0.003 0.002 0.003 0.002

Mean absolute error: 0.039 0.031 0.042 0.036

Minimum absolute

error: 0.001 0.001 0.001 0

Maximum absolute

error: 0.157 0.102 0.143 0.138

Pearson’s correlation

coefficient: 0.8048 0.7765 0.7220 0.7554

Percent within 5%: 43.363 45.545 39.823 46.018

Percent within 5% to 10%: 19.469 34.653 23.894 23.009

Percent within

10% to 20%: 30.973 14.851 29.204 28.319

Percent within 20% to 30%: 3.540 4.950 6.195 2.655

Percent over 30%: 2.655 0 0.885 0

Probability of correct

sign of increments

forecasting: 0,75 0,78 0,72 0,78

 

Hierarchical Extreme Learning Machine for

Unsupervised Representation Learning

Wentao Zhu and Jun Miao

Key Lab of Intelligent

Information Processing of

Chinese Academy of Sciences (CAS),

Institute of Computing Technology,

CAS, Beijing 100190, China

Email: wentao.zhu@vipl.ict.ac.cn;

jmiao@ict.ac.cn

Laiyun Qing

School of Computer and

Control Engineering,

University of Chinese Academy of Sciences,

Beijing 100049, China

Email: lyqing@ucas.ac.cn

Guang-Bin Huang

School of Electrical &

Electronic Engineering,

Nanyang Technological University,

Nanyang Avenue, Singapore 639798

Email: egbhuang@ntu.edu.sg

 

Abstract—Learning representations from massive unlabelled data is a topic for high-level tasks in many applications. The recent great improvements on benchmark data sets, which are achieved by increasingly complex unsupervised learning meth-ods and deep learning models with many parameters, usually requiring many tedious tricks and much expertise to tune. Additionally, the filters learned by these complex architectures are quite similar to standard hand-crafted visual features, and training to fine-tune the weights of deep architectures requires a long time. In this paper, the Extreme Learning Machine-Auto Encoder (ELM-AE) is employed as the learning unit to learn local receptive fields at each layer, and the lower layer responses are transferred to the last layer (trans-layer) to form a more complete representation to retain more information. In addition, some beneficial methods in deep learning architectures such as local contrast normalization and whitening are added to the implemented hierarchical Extreme Learning Machine networks to further boost the performance. The resulting trans-layer representations are processed into block histograms with binary hashing to produce translation and rotation invariant representations, which are utilized to do high-level tasks such as recognition and detection. The proposed trans-layer representa-tion method with ELM-AE based learning of local receptive filters was tested on the MNIST digit recognition data set, including MNIST variations, and on the Caltech 101 object recognition database. Compared to traditional deep learning methods, the proposed ELM-AE based system has a much faster learning speed and attains 65.97% accuracy on the Caltech 101 task (15 samples per class) and 99.45% on the standard MNIST data set.

I. INTRODUCTION

Almost all tasks such as classification, and recognition require us to design fine representations for their specific aims. For classification of images taken in natural settings, numer¬ous factors in the environment, such as lighting conditions, occlusions, corruptions and deformations, lead to large amount of intra-class variability. Good representations should reduce such non-informative intra-class variability, whilst preserving discriminative information across classes. However, designing good feature representations is a quite tough and difficult procedure for pattern recognition tasks [37].

This research is partially sponsored by Natural Science Foundation of China (Nos. 61272320 and 61472387).

 

Two basic approaches to research on feature representations are hand-crafted design and automated machine learning. For decades, enormous numbers of researchers and engineers have devised robust feature representations in their own domains. Many hand-crafted features are successfully proposed such as SIFT [1] and HoG [2] features in computer vision domain. However, these hand-crafted features have poor transferabil¬ity across domains. Novel features need to be redesigned elaborately when the domain of application is changed. The other way is machine learning of representations, which is a quite prevalent topic following the recent successes of deep learning [3]. Nevertheless, these fully learned representations by multilayer unsupervised learning followed by a fine-tuning procedure have too many parameters to be tuned, and require much expertise knowledge and sophisticated hardware.

In this paper, we implement a novel hierarchical neural network using the Extreme Learning Machine Auto Encoder (ELM-AE) [4] to very quickly learn local receptive filters and trans-layer representations. The hierarchical ELM trains the local receptive filters using ELM-AE without trivial BP fine-tuning, and the responses of the previous layer of our model are concatenated to that of the last layer to form a more complete representation. These trans-layer connections effec¬tively reduce the rapid information loss afflicting multi-layer unsupervised learning as commonly suffered by conventional NN architectures. In addition, local contrast normalization [5] and whitening are added in our hierarchical ELM network to boost its learning ability. The major difference between our hierarchical ELM network and conventional deep learning networks is that the hierarchical ELM network requires no back propagation information to fine-tune the feature banks with a trans-layer representation.

Experimental results indicate that the hierarchical ELM network boosts the ELM based methods effectively. The hierarchical ELM network is validated on digit recognition, object recognition tasks. Experimental results show that, the hierarchical ELM representations achieve accuracy of 6597% on Caltech 101 [7] 15 samples per class task and 9945% on MNIST data set [8].

We will start by reviewing related work on representation learning in Section II. Then the idea of the hierarchical

 

ELM network, including the pre-processing, ELM-AE learning local receptive filters, and trans-layer learning, is illustrated detailedly in Section III. How to use the hierarchical ELM network to extract features and tackle applications is also described in Section III. The experimental results and com-parative analysis on MNIST, MNIST variations, Caltech 101 data sets are presented on Section IV. Finally, discussion and conclusion are summarized in Section V.

II. RELATED WORKS

Much research has been conducted to pursuit a good representation by manually designing elaborative low-level features, such as SIFT feature [1], HOG feature [2] in computer vision field. However, these hand-crafted features cannot be easily adapted to new conditions and tasks, and redesigning them usually requires novel expertise knowledge and tough studies.

Learning representations is probably a promising way to handle the required elaborative expertise problem in devising hand-crafted features. Much recent work in machine learning has focused on how to learn good representations from massive unlabelled data, and great progresses have been made by deep neural networks [11]–[13]. The main idea of deep models is to learn multi-level features at different layers. High-level features generated in the upper-layer are expected to extract more complex and abstract semantics of data, and more invariance to intra-class variability, which is quite useful to high-level tasks. These deep learning methods typically learn multi-level features by greedily “pre-training” each layer using the specific unsupervised learning, and then fine-tuning the pre-trained features by stochastic gradient descent method with supervised information [3], [5]. However, these deep architectures have numerous parameters such as the number of features to learn, and parameters of unsupervised learning in each layer. Besides, the stochastic gradient descent also has various parameters such as momentum, weight decay rate, learning rate, and extra parameters including the Dropout rate or DropConnet rate in recently proposed convolutional deep neural networks (ConvNets) [11], [12].

There are also some works on conventional unsupervised learning methods with only single layer [10], [13]. The main idea of these methods is to learn over-complete representa¬tions with dense features. Although these methods have made progresses on benchmark data sets with almost no hyper parameters, these single layer unsupervised representational learning methods require over-complete features of dimensions as high as possible and the parameters need to be elaborately chosen in order to obtain satisfactory results [13].

A major drawback of deep learning methods with fine-tuning for learning representations is the big consumption of expensive computational resources and high complexity of the models. One intuition is that, since the elaborately learned features are quite similar to some conventional un¬supervised features, such as wavelets and PCA, why not jump over the tough and time-consuming fine-tuning procedure and take those features stacked directly. Furthermore, more robust invariant features can be better devised other than various pooling strategies. Wavelet scattering networks (ScatNet) are such networks with pre-fixed wavelet filter banks in the deep 

 

convolutional architectures [14]. The ScatNets have quite solid mathematical analysis of their rotation and translation invariants at any scale. More surprisingly, superior perfor¬mance over ConvNet and deep neural networks is obtained by the ScatNets’ pre-fixed feature bank networks with no fine-tuning phase. However, the ScatNet is shown to have inferior performance in large intra-class variability including great illumination changing and corruption [15].

The other non-propagation deep network with pre-fixed feature banks is PCANet [15]. The PCANet uses two layer cascaded linear networks with na¨ıve PCA filter banks to extract more complex features. The PCANet also utilizes the histogram operator popular in computer vision domain to retain translation and rotation invariance [1], [2]. The output of the two layer cascaded PCA network is processed by the quantized histogram units. The PCANet presents a superior or highly comparable performance over other methods such as ScatNet [14], ConvNet [11] and HSC [16], especially in face recognition tasks with large occlusion, illumination, expression and pose changes. However, the cascaded PCA structure in PCANet will face great information loss and corruption after multilayer transformation. Besides, the learning ability of na¨ıve PCA unsupervised learning is much weaker than other learning units.

On the other hand, Extreme Learning Machine (ELM) is a popular method famous for its fast learning speed and good generalization ability [17], and has been used in many applications [18], [19]. Many methods have been proposed to further boost its performance [20]–[22]. Also, ELM has been extended to multilayer neural networks [4], [23]. However, the accuracies of these methods are lower than the current deep learning methods.

This paper demonstrates a novel method, the hierarchical ELM network, using the Extreme Learning Machine (ELM) to learn the local receptive fields fast without trivial BP fine-tuning. Moreover, the hierarchical ELM network combines the first layer representation with the second layer representation. Furthermore, the local contrast normalization and whitening operators are applied to ameliorate the ELM filter learning in the hierarchical ELM network. Thus the implemented network can form a more complete and efficient representation, whilst retaining the advantages such as faster learning speed and more reliable performance. Also, the successfully used histogram operator in computer vision domain is adopted to preserve translation and rotation invariance after binary hashing quanti-zation. Experimental results confirm that the hierarchical ELM network learns effective feature representations that contribute to object recognition and digital recognition tasks.

III. HIERARCHICAL EXTREME LEARNING MACHINE

In this section, we present a novel framework, the hierar-chical ELM network, for unsupervised representation learning, illustrated in Fig. 1. The procedures of the hierarchical ELM network is similar to other commonly used frameworks in computer vision [15], [24] and other feature learning work [13] as well. Different from the traditional methods, the hierarchical ELM network utilizes the unsupervised Extreme Learning Machine-Auto Encoder (ELM-AE) to learn the local receptive features, and needs no fine-tuning procedure to adjust

 

those local filters. Besides, the previous layer’s unsupervised features transformed by ELM-AE local receptive features are concatenated to the last layer to form a much more completed representation, which is shown quite effective for the following tasks.

Applications: Classification, recognition, -- 

Fig. 1: The framework of the hierarchical ELM network. The hierarchical ELM includes three layer, the first layer of ELM-AE local receptive mapping, the second layer of ELM-AE local receptive mapping, and the output layer with previous responses inputs from previous layers.

In the beginning of each layer (the first and second layer), the system extracts large numbers of random sub-patches from unlabelled input images. Each patch has dimension k1-by-k2, called receptive field size. (Suppose the images used here are all gray images. Actually, if the images are colored images with d channels, the patch dimension is k1-by-k2-by-d. Just process the other d  1 channels the same as following procedures step by step independently.) Then a data set of m patches is constructed, X = x(1), •••, x(), where x() R12 stands for the ith patch extracted from the input images in the first layer or the input feature maps in the second layer. Given the data set before the processing of the first and second layer, we apply the preprocessing of local contrast normalization (LCN) and whitening, and Extreme Learning Machine-Auto Encoder (ELM-AE) unsupervised learning of each layer to learn local receptive fields in the first and second layer. 

 

A. Pre-processing-Local Contrast Normalization and Whiten¬ing

The first pre-processing method of each ELM-AE layer is local contrast normalization (LCN) [5]. For each local patch x() in the extracted patch data set X, we normalize the patch x() by subtracting its mean and dividing by its standard deviation as,

y() = (x() 1 

12

VI~

1: 1 1:2 1: 1 1:2

( 1 (x() 1 x())2 + C),

12 12

=1 =1 =1 =1

j= 1,••• , k1; k = 1,••• , k2; i = 1, ••• , m,

where C is a constant integer to make the model more robust commonly set in practice. In the work, we set C as 10.

( i)

Output

T

P1T PL

1 ... L

(WL, bL)

(W1, b1)

k1k2

( i)

Fig. 2: The framework of ELM-AE. The ELM-AE is a network using ELM to reconstruct the input.

The LCN has explicit explanations both in physics and physiology. The mean of local patch stands for local bright¬ness, and the standard deviation represents contrast normaliza¬tion. By LCN, the illumination and material optimal property effects of the processed patches are removed. On the other hand, the LCN has an effect similar to lateral inhibition found in real neurons. The LCN operator inhibits the responses within each local patch, whilst activating responses in the same location of these patches.

Following LCN, whitening is the second preprocessing method for each ELM-AE unsupervised layer. Whitening is commonly used in various applications and a decorrelation operation, which reduces redundant representation of images. Whitening is commonly used in various applications and we

 

do not describe it in details. The whitening operator transforms the patches as,

[D U]=  ( (Y))

z() = U( + ())12U y() = 1•• • (2)

where Y is formed by  patches y(), () stands for covariance function and the size of output data is 1 * 2, () is the eigenvalue decomposition function, D and U are eigenvalues and eigenvectors respectively,  is set as 01 here. The Zeros Component Analysis (ZCA) whitening also has biological explanation and has been proved its effectiveness by a lot of work.

B. Extreme Learning Machine-Auto Encoder to Learn Local Receptive Filters in Each Layer

In the hierarchical ELM network, we use a novel approach, Extreme Learning Machine-Auto Encoder (ELM-AE), to learn local receptive filters in each layer, illustrated in Fig. 2.

Assuming that the number of first layer feature banks is 1, flatten each pre-processed patch z() extracted from input images, and put the flattened vectors together. The matrix formed by the extracted  patches will be obtained as

Z1= I z1(1) z1(2) ••• z1() I C R12 (3)

The objective function of ELM-AE is

11Z1  2 (1(WZ1 + bi))1122 + 111122

 WW = I1  b b = 1

where W is the orthogonally random weights with size of 1x 12, 1 is the number of local receptive filters in the first layer, I1 is the identity matrix of size 1, b is the normalized random vector of dimensions 1, i is the column vector of size  and full of the element 1 to make b a matrix, () is the element wise sigmoid function, 1 and 2 are the factors changing the hidden nodes and output nodes distribution of ELM-AE based on intrinsic plasticity property respectively, is the output layer weight, C is the trade-off between errors and model complexity. Then the output weight  is calculated as [4]

The first layer local receptive filters are 1 • • •L1. For each input image, the first layer L1 feature maps can be obtained by

I1() = I * ii = 1 •••L1  (6)

where I stands for an input image with zero padded to make I1() has the same size as the input image, and * stands for two dimensional convolution operator.

Patches should be extracted from feature maps I1() ob¬tained from the first ELM-AE layer. These patches are also transformed by the pre-processing LCN and whitening op-erators. Then applying the same process as the first layer, we obtain the second layer feature map for each first layer feature map, I2() = 1 • • •2  where 2 is the number of filters in the second layer.

 

For an input image, we get 1 x (2 + 1) feature maps

after the two ELM convolution layers by concatenating the

first layer maps to the second layer. That is

  

I1 (1) I1 (2) • • •  I1 (1) I21  I2

(1) 1  ••• 

(2)

I21  •••  I2

(2) 1  I2

(1) 1  •••  I2

(2) (2)  (7)

1

where I1() stands for the th feature map in the first layer, and I2  stands for the th feature map of the th first layer ()

feature map in the second layer.

C. Block-wise Histogram

Fig. 3: The output layer of the hierarchical ELM network. Firstly, quantize the responses from previous layers into binary values. Secondly, compress these responses. Thirdly, use the histogram to retain the invariance.

The output layer of the hierarchical ELM network is illustrated in Fig. 3. For an input training image, the first step is to use binary hashing to make the constructed real valued

1 x (2 + 1) feature maps have only two values, 0 and 1. The second step is to compress these binary feature maps by quantizing each 1 binary feature maps. The number of second layer filters 1 is set as 8, and the number of first layer filters

2 is set the same as 1. That is, we compress each 1 binary feature maps into one feature map, and the compressed feature maps have pixel values from 0 to 255 [15]. Then we get 2+1 compressed feature maps for each training image as

(

2+1)1

I (1) I (2) • • •  I(2) I  (8)

The third step is to construct block-wise histogram il¬lustrated as the third procedure in Fig. 3. First, we should partition each encoded feature map. Assuming that the size of

 

compressed map is x x y, and the size of block is w1 x w2 with strides s1xs2, each compressed feature map is partitioned into L(x  w1)/s1 + 1j x L(y  w2)/s2 + 1j blocks. For all the L2 + 1 compressed feature maps of each input image, we get blocks as

{B1, B2, ••• , B(L2+1)L(xw1)/s1+1JL(yw2)/s2+1f ~ , (9)

where Bi stands for the ith blocks constructed from the compressed feature maps. Next step is to build histograms in each of the blocks. In the hierarchical ELM network, we set the number of bins in the histogram as 2L1. It means that each integer of the pixel values is set as a bin and a sparse vector representing the histogram is constructed. Then concatenate these NB histograms to form the completed representation of the input image as

f(Image) = [hist(B1)T , hist(B2)T , •••,hist(BNB )T ]T E RNB(2L1 ), (10)

where hist()stands for histogram operator, and NB = (L2 +

1) x L(x  w1)/s1 + 1j x L(y  w2)/s2 + 1j. Then we use the hierarchical ELM network representation of each training images to learn a dimension reduction weight, or to train a classifier to tackle the next applications directly.

IV. PERFORMANCE EVALUATION

In this section, we will introduce how to train the hierar-chical ELM network, including how to select the parameters. The hierarchical ELM networks are validated on MNIST [8], MNIST variations [9] and Caltech 101 [7] data sets. MNIST and MNIST variation data sets are grey hand-written digits of ten classes and size 28 x 28. Other description of the data sets are illustrated as the Table I. Caltech 101 data set contains color images belonging to 102 categories including a background class. The number of each class’s images varies from 31 to 800.

TABLE I: Data Set Description about the Size of Training, Validation and Testing.

Data Sets Training Validation Testing

MNIST 50,000 10,000 10,000

MNIST basic 10,000 2,000 50,000

MNIST Rotation 10,000 2,000 50,000

MNIST back-rand 10,000 2,000 50,000

MNIST back-image 10,000 2,000 50,000

MNIST back-image-rotation 10,000 2,000 50,000

Convex vs. non-convex 6,000 2,000 50,000

A. How to Train

In the hierarchical ELM network, there are mainly two kinds of parameters, local receptive fields related parameters and ELM-AEs related parameters.

For the local receptive fields related parameters such as the block size, the stride size and the patch size are determined through cross-validation. On all the tasks, the filter size is set to 7 x 7 pixels, the number of filters is set to L1 = L2 = 8 and the size of strides is the half size of the block. For MNIST, MNIST basic, and mnist-rotation data sets, the block 

 

size is 7 x 7 pixels. For mnist-back-rand, mnist-back-image and mnist-rot-back-image data sets, the block size is 4 x 4 pixels. For convex data set, the block size is 28 x 28. For Caltech-101 data set, the block size is set to a quarter of the image size, and the whitening PCA (WPCA) is used to reduce the dimension of each block representation from 256 to 64. A simple linear SVM with default parameters C = 1 is connected to the hierarchical ELM representation to do recognition with much fast learning speed [26].

For the ELM-AEs related parameters such as C, α1, α2 in each layer, we selected these parameters by layer-wise greedy validation on the validation set. We first selected the parameter

C, then fixed it and selected α1, α2. α1, α2 were selected in a decreasing order 1, 0.9, 0.8, • • •, 0.5. For MNIST data set, we used the parameters the same as the MNIST basic data set. The training set of Caltech 101 was constructed as randomly selected 15 samples for each class, and the validation set of Caltech 101 was constructed the same as training set. The parameters of each data set are listed in the Table II.

TABLE II: Parameters of ELM-AEs on Each Data Set

Data Sets The first layer The second layer

α1 α2 C α1 α2 C

Basic 0.9 0.7 106 1 0.6 106

Rot. 1 0.7 105 0.8 1 105

Back-rand 0.7 0.9 102 0.7 1 107

Back-im. 0.8 0.5 108 0.7 0.7 108

Back-im.-rot. 1 0.7 103 1 1 107

Conv. 1 1 104 1 0.9 106

Caltech-101 0.5 1 102 0.6 0.9 107

B. Effect of Whitening

We validated the whitening effect of ELM-AE filter learn-ing. Whitening is an effective and helpful pre-processing procedure for many methods such as PCA and conventional deep models. However, the effect of combining whitening with ELM-AE has not been validated. The parameters of the hierarchical ELM network without whitening or with whitening are both selected as the same procedure as the above description respectively. The performances of the hierarchical ELM network with whitening pre-processing and without pre¬processing on MNIST and MNIST variation data sets are listed in Table III.

From Table III, we observe that whitening pre-processing boosts the hierarchical ELM networks. That is to say, whiten¬ing is also an effective pre-processing method in ELM-AE filter learning. We also printed the filter maps of the hierarchi¬cal ELM networks on the seven data sets. The ELM-AE filters with whitening and without whitening are illustrated from Fig. 4. to Fig. 9.

From these figures, whitening seems to alter the crispness of the images. However, from the performance view, ELM-AEs with whitening have better performance. The seeming ironical results bring two possible causes. The first is whether the seeming better filters are the true better filters in the hierarchical models, at least in the hierarchical ELM networks. The second is the ELM-AEs cannot learn the representation

 

of small patches robustly since the compressive representation loses lots of information in ELM-AE. We will further investi-gate the problem in the future and it seems a quite promising question.

C. Digital Recognition on MNIST and MNIST Variations Data Sets

We tested our implemented hierarchical ELM network on the MNIST, MNIST variations data sets as the protocols of these data sets. The results are listed in Table IV and Table V.

TABLE III: The Error Rates on MNIST Data Set. The Perfor-mance of Hierarchical ELM Is 9945% with A Fast Speed. Note That The Performances Are Only Those of Methods without Using Augmented Samples.

Method Error Rate (%) Training Time

SDAE [35] 1.28 17 hours

SAE [35] 1.4 17 hours

DBM [36] 0.95 19 hours

DBN [36] 1.13 5.7 hours

ML-ELM [4] 0.97 7.5 mins

HSC [16] 0.77 -

ConvNet [5] 0.53 -

Conv. Maxout + Dropout [27] 0.45 -

ScatNet (rbf ) [14] 0.43 -

PCANet [15] 0.66 -

Hierarchical ELM 0.55 1.5 hours

From Table IV, the results show that the hierarchical ELM network is only inferior to ScatNet-2 and Convolution Network related methods with enhanced tricks. It is worthy to mention that the performance of ScatNet-2 is achieved by being con-nected with a non-linear SVM with RBF kernels with tuned parameters, but our model is connected with linear SVM with all default parameters ( = 1). Our models performance (0.55) is highly close to that of ConvNet (0.53) on MNIST dataset [5]. Also, the learning speed of the hierarchical ELM network is fast, 1.5 hours to get the representations of all training samples on Intel(R) Core(TM) i7-3770 CPU @ 3.4GHz 32GB RAM, and it boosts the performances of ELM based learning methods greatly.

TABLE IV: The Error Rates (%) of Hierarchical ELM Network with Whitening and without Whitening. The Performances using Whitening Are Better than Those without Whitening Consistently.

Data Sets Without whitening With whitening

MNIST 0.78 0.55

MNIST basic 1.05 1

MNIST Rotatation 7.09 7.02

MNIST back-rand 6.17 6.13

MNIST back-image 11.54 9.89

MNIST back-image-rotation 39.03 34.45

Convex vs. non-covex 5.88 3.81

From Table V, we observe that the hierarchical ELM network achieves the best performance on four of six data

The matlab code can be downloaded from https://github.com/wentaozhu/ Hierarchical-ELM-Network.git

 

TABLE V: The Error Rates (%) on MNIST Variations Data Sets. The Performance of Hierarchical ELM Is The Best on Average.

Method basic rot bk-rand bk-im bk-im-rot conv

CAE-2 [28] 2.48 9.66 10.9 15.5 45.23 -

TIRBM [29] - 4.2 - - 35.5 -

PGBM+DN-1 [30] - - 6.08 12.25 36.76 -

ScatNet [14] 1.27 7.48 18.4 12.3 50.48 6.5

PCANet [15] 1.06 7.37 6.19 10.95 35.48 4.36

Hierarchical ELM 1 7.02 6.13 9.89 34.45 3.81

sets with a simple linear SVM classifier. It is sufficient to indicate that the hierarchical ELM network works well in the convolutional structure with ELM-AE local receptive filters and trans-layer structure.

D. Object Recognition on Caltech 101 Data Set

The pre-processing of the data set is to convert the images into grey scale, and adjust the longer side of the image to 300 with preserved aspect ratio. Two tasks are conducted. One is with a training set of 15 samples per class. The other is with a training set of 30 samples per class. The training sets are randomly sampled from Caltech 101, and the rest are testing set. Five rounds of experiments are recorded, and the performance is recorded as the average of the five rounds of results. The results are listed in Table VI.

TABLE VI: The testing Accuracy Rates (%) on Grey Level Images of Caltech 101 Data Set. The Hierarchical ELM Achieves 6597% and 7374% on 15 samples per class and 30 samples per class respectively.

Methods 15 samples per class 30 samples per class

CDBN [25] 57.70  1.50 65.40 0.50

ConvNet [31] 57.60  0.40 66.30 1.50

DeconvNet [32] 58.60  0.70 66.90 1.10

Chen et al. [33] 58.20  1.20 65.80 0.60

Zou et al. [34] -66.5

HSC [16] -74.0

PCANet [15] 61.46  0.76 68.56 1.01

Hierarchical ELM 65.97  0.69 73.74  0.37

The Table VI shows that the hierarchical ELM network gets the accuracy of 6597% trained by 15 samples per class and 7374 on 30 samples per class tasks respectively. The demonstrated network with ELM local receptive filters really makes a high progress for the data set and ELM based learning.

V. CONCLUSION AND DISCUSSION

In this paper, a novel unsupervised representation model, the hierarchical ELM network, is implemented. The method learns the local receptive filters by ELM-AE without trivial BP fine-tuning. Several key elements, such as LCN, whitening and trans-layer representation are combined together to boost the performance. We also provide a training procedure of ELM-AE in the hierarchical network. The hierarchical ELM network achieves accuracy of 6597% on Caltech 101 15 samples per

 

class task and 9945% on MNIST data set with relatively fast learning speed.

However, as presented in the experiment section, the hier¬archical ELM network cannot robustly learn the visually fine local receptive filters. There seems a promising research topic in the future to boost the ELM-AE learning ability on filters. Also, learning filters by ELM-AE on other convolutional architectures is a promising topic. And the hierarchical ELM is just a start. There is a lot of room for great improvements to come from further extensions of the hierarchical ELM.

Fig. 4: Local receptive filters of 2 layers’ ELM-AE on MNIST basic data set with whitening (up 16) vs. without whitening (bottom 16)

 

Fig. 5: Local receptive filters of 2 layers’ ELM-AE on MNIST rotation data set with whitening (up 16) vs. without whitening (bottom 16)

 

 

 

Fig. 7: Local receptive filters of 2 layers’ ELM-AE on MNIST back-im data set with whitening (up 16) vs. without whitening (bottom 16)

Fig. 8: Local receptive filters of 2 layers’ ELM-AE on MNIST back-im-rot. data set with whitening (up 16) vs. without whitening (bottom 16)

[2] N. Dalal, and B. Triggs, “Histograms of oriented gradients for human detection,” In IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2005), June 2005, Vol. 1, pp. 886-893, IEEE.

[3] G. E. Hinton, and R. R. Salakhutdinov, “Reducing the dimensionality of data with neural networks,” Science, vol. 313(5786), pp. 504-507, 2006.

[4] L. L. C. Kasun, H. Zhou, G.-B. Huang, and C. M. Vong, “Represen-tational learning with extreme learning machine for big data,” IEEE Intelligent Systems, vol. 28(6), pp. 31-34, December 2013.

[5] K. Jarrett, K. Kavukcuoglu, M. Ranzato, and Y. LeCun, “What is the best multi-stage architecture for object recognition?,” In Proceedings of IEEE 12th International Conference on Computer Vision (ICCV 2009), pp. 2146-2153, September 2009.

[6] A. Coates, A. Y. Ng, and H. Lee, “An analysis of single-layer networks in unsupervised feature learning,” In Proceedings of International Con¬ference on Artificial Intelligence and Statistics, pp. 215-223, 2011.

[7] L. Fei-Fei, R. Fergus, and P. Perona, “Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories,” Computer Vision and Image Understanding, vol. 106(1), pp. 59-70, 2007.

 

 

Fig. 6: Local receptive filters of 2 layers’ ELM-AE on MNIST back-rand data set with whitening (up 16) vs. without whiten¬ing (bottom 16)

REFERENCES

[1] D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” International journal of computer vision, vol. 60(2), pp. 91-110, 2004.

 

Fig. 9: Local receptive filters of 2 layers’ ELM-AE on Convex data set with whitening (up 16) vs. without whitening (bottom 16)

 

[8] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” In Proceedings of the IEEE, vol. 86(11), pp. 2278-2324, 1998.

[9] H. Larochelle, D. Erhan, A. Courville, J. Bergstra, and Y. Bengio, “An empirical evaluation of deep architectures on problems with many factors of variation,” In Proceedings of the 24th international conference on Machine learning, pp. 473-480, ACM, June 2007.

[10] Q. V. Le, A. Karpenko, J. Ngiam, and A. Y. Ng, “ICA with recon-struction cost for efficient overcomplete feature learning,” In Proceedings of Advances in Neural Information Processing Systems, pp. 1017-1025, 2011.

[11] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” In Proceedings of Advances in neural information processing systems, pp. 1097-1105, 2012.

[12] L. Wan, M. Zeiler, S. Zhang, Y. L. Cun, and R. Fergus, “Regularization of neural networks using dropconnect,” In Proceedings of the 30th International Conference on Machine Learning (ICML-13), pp. 1058¬1066,2013.

[13] A. Coates, and A. Y. Ng, “Learning feature representations with k-means,” In Neural Networks: Tricks of the Trade, Lecture Notes in Computer Science, vol. 7700, pp. 561-580, Springer Berlin Heidelberg, 2012.

[14] J. Bruna, and S. Mallat, “Invariant scattering convolution networks,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 35(8), pp. 1872-1886, 2013.

[15] T. H. Chan, K. Jia, S. Gao, J. Lu, Z. Zeng, and Y. Ma, “PCANet: A Simple Deep Learning Baseline for Image Classification?,” arXiv preprint arXiv:1404.3606, 2014.

[16] K. Yu, Y. Lin, and J. Lafferty, “Learning image representations from the pixel level via hierarchical sparse coding,” In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2011, pp. 1713-1720.

[17] G.-B. Huang, Q.-Y. Zhu, and C.-K. Siew, “Extreme learning machine: a new learning scheme of feedforward neural networks,” In Proceedings of International Joint Conference on Neural Networks, vol.2, pp. 985-990, July 2004.

[18] W. Zhu, J. Miao, J. Hu, and L. Qing, “Vehicle detection in driving simulation using extreme learning machine,” Neurocomputing, vol. 128, pp. 160-165, 2014.

[19] Q. Leng, H. Qi, J. Miao, W. Zhu, and G. Su, “One-Class classification with extreme learning machine,” Mathematical Problems in Engineering, in press.

[20] W. Zhu, J. Miao, and L. Qing, “Constrained extreme learning machine: a novel highly discriminative random feedforward neural network,” International Joint Conference on Neural Networks (IJCNN), pp. 800¬807,2014.

[21] W. Zhu, J. Miao, and L. Qing, “Robust regression with extreme support vectors,” Pattern Recognition Letters, vol. 45, pp. 205-210, 2014.

[22] W. Zhu, J. Miao, and L. Qing, “Extreme support vector regression,” In Extreme Learning Machines 2013: Algorithms and Applications, pp. 25-34, Springer International Publishing, 2014.

[23] Z. Bai, G.-B. Huang, L. L. C. Kasun, C. M. Vong, and C. L . P. Chen, “Local receptive fields based extreme learning machine,” IEEE Computational Intelligence Magazine, vol. 10(2), 2015.

[24] Y.-L. Boureau, F. Bach, Y. LeCun, and J. Ponce, “Learning mid¬level features for recognition,” In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 13-18 June 2010, pp. 2559-2566.

[25] H. Lee, R. Grosse, R. Ranganath, and A. Y. Ng, “Convolutional deep belief networks for scalable unsupervised learning of hierarchical repre¬sentations,” In Proceedings of the 26th Annual International Conference on Machine Learning, pp. 609-616, ACM, June 2009.

[26] R. E. Fan, K. W. Chang, C. J. Hsieh, X. R. Wang, and C. J. Lin, “LIBLINEAR: A library for large linear classification,” The Journal of Machine Learning Research, vol. 9, pp. 1871-1874, 2008.

[27] I. J. Goodfellow, D. Warde-Farley, M. Mirza, A. Courville, and Y. Ben-gio, “Maxout networks,” arXiv preprint arXiv:1302.4389, 2013.

[28] S. Rifai, P. Vincent, X. Muller, X. Glorot, and Y. Bengio, “Contractive auto-encoders: Explicit invariance during feature extraction,” In Proceed¬ 

 

ings of the 28th International Conference on Machine Learning (ICML-11), pp. 833-840, 2011.

[29] K. Sohn, and H. Lee, “Learning invariant representations with local transformations,” arXiv preprint arXiv:1206.6418, 2012.

[30] K. Sohn, G. Zhou, C. Lee, and H. Lee, “Learning and Selecting Features Jointly with Point-wise Gated B oltzmann Machines,” In Proceedings of The 30th International Conference on Machine Learning, pp. 217-225, 2013.

[31] K. Kavukcuoglu, P. Sermanet, Y. L. Boureau, K. Gregor, M. Mathieu, and Y. L. Cun, “Learning convolutional feature hierarchies for visual recognition,” In Proceedings of Advances in neural information process¬ing systems, pp. 1090-1098, 2010.

[32] M. D. Zeiler, D. Krishnan, G. W. Taylor, and R. Fergus, “Deconvo-lutional networks,” In Proceedings of IEEE Conferences on Computer Vision and Pattern Recognition (CVPR), June 2010, pp. 2528-2535.

[33] B. Chen, G. Polatkan, G. Sapiro, L. Carin, and D. B. Dunson, “The hierarchical beta process for convolutional factor analysis and deep learn¬ing,” In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pp. 361-368, 2011.

[34] W. Zou, S. Zhu, K. Yu, and A. Y. Ng, “Deep learning of invariant features via simulated fixations in video,” In Proceedings of Advances in Neural Information Processing Systems, pp. 3212-3220, 2012.

[35] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P. A. Manzagol, “Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion,” The Journal of Machine Learning Research, vol. 11, pp. 3371-3408, 2010.

[36] R. Salakhutdinov, and H. Larochelle, “Efficient learning of deep boltz-mann machines,” The Journal of Machine Learning Research, vol. 9, pp. 693-700, 2010.

[37] B. Widrow, and J. C. Aragon, “Cognitive memory,” Neural Networks, vol. 41, pp. 3-14, 2013.

 

 

Sparse Recovery Using Sparse Matrices

 

The MIT Faculty has made this article openly available. Please share how this access benefits you. Your story matters.

Citation Gilbert, Anna, and Piotr Indyk. “Sparse Recovery Using Sparse Matrices.” Proceedings of the IEEE 98.6 (2010): 937–947. Web. ©2010 IEEE.

As Published http://dx.doi.org/10.1109/jproc.2010.2045092

Publisher Institute of Electrical and Electronics Engineers

Version Final published version

Accessed Sat Dec 02 23:35:32 EST 2017

Citable Link http://hdl.handle.net/1721.1/70932

Terms of Use Article is made available in accordance with the publisher's policy and may be subject to US copyright law. Please refer to the publisher's site for terms of use.

Detailed Terms

 

 

 

 

INVITED

PAPER

Sparse Recovery Using

Sparse Matrices

Significant results in predicting the operation of equipment such as network routers,

or the results of group testing for defective items, can often be obtained

from a few samples.

By ANNA GILBERT AND PIOTR INDYK

 

ABSTRACT | In this paper, we survey algorithms for sparse recovery problems that are based on sparse random matrices. Such matrices has several attractive properties: they support algorithms with low computational complexity, and make it easy to perform incremental updates to signals. We discuss applications to several areas, including compressive sensing, data stream computing, and group testing.

KEYWORDS | Compressive sensing; expanders; sparse matrices; sparse recovery; streaming algorithms

I. INTRODUCTION

The past several years have seen a new approach to the acquisition of compressible signals. Traditional approaches first capture the entire signal and then process it for compression, transmission, or storage. In comparison, the new approach obtains a succinct approximate representation directly by acquiring a small number of nonadaptive linear measurements of the signal. For any signal x, of length n, the representation is equal to Ax, where A is an m x n matrix. The vector Ax is often referred to as the measurement vector or sketch of x. Although m is typically much smaller than n, the sketch Ax contains plenty of useful information about the signal x. In particular, the sketch of x retains enough inherent information that we can directly obtain a sparse approximation or compressed form of the signal.

Manuscript received April 20, 2009; revised November 22, 2009; accepted February 26, 2010. Date of current version May 19, 2010. This work was supported by a David and Lucille Packard Fellowship, the MADALGO (Center for Massive Data Algorithmics, funded by the Danish National Research Association), and the National Science Foundation (NSF) under Grants CCF-0728645, CCF-0910765, and DMS-0547744.

A. Gilbert is with the Department of Mathematics, University of Michigan, Ann Arbor, MI 48109-0937 USA (e-mail: annacg@umich.edu).

P. Indyk is with the Department of Electrical Engineering and Computer Science, MIT, Cambridge, MA 02139-4307 USA (e-mail: indyk@mit.edu).

Digital Object Identifier: 10.1109/JPROC.2010.2045092

0018-9219/$2 6. 00 @2010 IEEE

 

This approach has been discovered and explored exten-sively in several different research communities, including theoretical computer science, applied mathematics, and digital signal processing. The goal of that research is to obtain encoding and recovery schemes with good compression rate (i.e., short sketch lengths) as well as good algorithmic properties (i.e., low encoding, update and recovery times).

Linear sketches have found numerous uses in several areas, including compressive sensing, data stream computing, and combinatorial group testing.

• Compressive sensing. In this area [12], [16], the signal or image x is acquired using (analog or digital) hardware, which (approximately) com¬putes a dot product of each row of the matrix A and the signal at a unit cost. Once we obtain the measurement vector Ax, we process it digitally to extract information about the signal, including significant coefficients in an orthonormal basis (e.g., wavelet or Fourier), as well as the original signal. Frequently, the number of measurements we obtain with compressed sensing hardware is much less than that of traditional hardware devices. Nevertheless, we can sample a bandlimited analog signal at a sub-Nyquist rate and still recover significant frequencies in the signal or the entire signal spectrum. For examples of compressive sensing hardware, see [13], [31], [41], and [42].

• Data stream computing. In this area [7], [24], [33], the vectors x are often very large, and cannot be represented explicitly. For example, in network measurement, xi could denote the total number of packets with destination i passing through a network router. Storing such vector x itself is typically infeasible due to its large size: each destination is represented by an IP address that is 32-bit long, and therefore the vector x has dimension n 1/4 232. Thus, it is preferable to maintain a lower dimensional sketch

Vol. 98, No. 6, June 2010 | PROCEEDINGS OF THE IEEE 937

 

Gilbert and Indyk: Sparse Recovery Using Sparse Matrices

 

Ax instead and recover an approximation to x from the sketch. However, it must be possible to maintain such sketch under incremental updates to x. For example, if a new packet arrives, the corresponding coordinate of x is incremented by 1, which should be reflected in the sketch Ax. Fortunately, this can be easily done if the sketching procedure is linear. Specifically, let Ai denote the update to the vector x after seeing a packet with destination i (i.e., Ai = 1 and 0j = 0 for j =6 i). Then, we have A(x + Di) = Ax + ADi. Since ADi is simply the ith column of A, updating the sketch can be accomplished by simply adding that column to the current sketch Ax. See, e.g., [18] and [30] for more information about using data stream algorithms for network measurement.

• Combinatorial group testing. In pooling designs or more generally combinatorial group testing [15], the vector x represents a universe of n items in total. Moreover, we know k of the elements are defective. More specifically, the vector x is the characteristic vector for the defective set so that x E {0, 1}n has exactly k entries that are 1 and (n - k) zeros. The goal of combinatorial group testing is to construct a collection of tests (called a design) to minimize the number of tests needed to find the defective set for the worst case input. The tests are represented by a matrix A that is binary, with the jth column of the ith rows equal to 1 if and only if the jth item is used by the ith test. In the simplest (boolean) setting, each test returns 1 if at least one of the elements used in the test is defective. In our setting we assume the linear model, where each test returns the number of defec¬tive elements. Note that each such test corresponds to taking the dot product of x and a test vector, and therefore is captured in the linear sketching model.1 See, e.g., [17], [28], and [36] for further discussion and recent developments in the area.

In each of these applications, it is useful (and often crucial) that the measurement matrix A be a sparse matrix, i.e., contain very few nonzero elements per column. In data stream processing, the time needed to update the sketch Ax under the update Ai is proportional to the number of nonzero elements in the vector ADi, which is equal to the number of nonzeros in the ith column of A. In experiment pooling, the design matrix A is a binary matrix that captures which compounds are pooled together and the measure¬ments Ax reflect the activity levels of the pooled compounds. In many chemical and biological applications, the assumption that compound activity is a linear function of the concentra¬tion holds only when there are not many compounds mixed together in a single pool; thus, the design matrix A should be

1In fact, we can assume an even more general setting, where we allow a general vector x E Rn, and our goal is to identify the top k most significant coefficients from the set of linear measurements. This is applicable in a setting where the entries in x represent the activity level of n compounds, or a genetic response in a biological sample.

938 PROCEEDINGS OF THE IEEE | Vol. 98, No. 6, June 2010

 

not only binary but also sparse. In other applications, sparsity can be useful for computational reasons: one can compute the matrix–vector product Ax very quickly.2

proximation error IIx - x'IIq, where x' ranges over all k-sparse vectors (i.e., that have at most k nonzero entries). Our goal is, given Ax, to find a vector x^ such that the 2p approximation error3 IIx - ^xkp is at most c > 0 times Errkq (x), i.e.,

II^x - xIIp < c • Errkq(x). (1)

Note that for any value of p, the error IIx - ^xIIp is minimized when x^ consists of the k largest (in magnitude) coefficients of x. We refer to such x^ as the "head" of the signal x, while x - x^ will be called the "tail" of x.

As mentioned earlier, we aim to design sparse recovery schemes that achieve short sketches, have low algorithmic complexity, and provide "good" recovery guarantees. In addition, the schemes described in this survey can be classified based on other characteristics, such as 1) whether the schemes are randomized or deterministic, or 2) how general is the class of signals x supported by the schemes. In the following, we elaborate on both issues.

• Randomization: we distinguish between two clas¬ses of schemes: for-each and for-all. The latter describes a scheme in which one matrix A works for all signals x. In the former case, the matrix A is chosen at random from some distribution, and for each signal x, the recovery algorithm works "with high probability" (at least 1 - 1/n).4 Naturally, schemes with the for-all property are preferable to those with the for-each guarantee (if all other parameters are the same).We note that "for-all" does not mean that the matrix is constructed in an "explicit" or efficient manner. In fact, most of the constructions presented here use the probabilistic method. Although it is possible to construct recovery schemes explicitly [2], [14], [34], such schemes tend to require more measurements.

2Specifically, the matrix–vector product can be computed in time O(ns), where s is the column sparsity of A. As we will see in Section III, in many settings, one can achieve s = O(log(n/k)), which leads to the running time of O(n log(n/k)). This compares favorably to the O(nm) time achievable for random Gaussian matrices, or to the O(nlogn) time achievable for random Fourier-like matrices.

3It is natural to consider p = q. However, as we will see later, other guarantees are also possible.

4We adopt here the terminology frequently used in computer science. Note that one could require weaker probability bounds, e.g., 1- o(1). However, all algorithms presented in this survey naturally achieve the stronger probability bound without changing the (asymptotic) bound on the number of measurements.

 

Gilbert and Indyk: Sparse Recovery Using Sparse Matrices

 

• Generality of supported signals: ideally, the recovery schemes should support arbitrary signals x. In this survey, we focus on describing such schemes. However, there has been plenty of work on algorithms supporting more restrictive classes of signals. In particular, there have been several schemes based on sparse matrices that work for (almost) exactly k-sparse signals [27], [29], [32], [37]–[39], [43], [44]. Although we do not cover them in detail, we point out relevant connections and references whenever possible.

B. Survey Summary

We present an overview of the algorithms for sparse recovery that utilize sparse measurement matrices. The description is divided into two sections: for-each algo¬rithms are covered in Section II, while for-all algorithms are described in Section III. Historically, most of the for-each schemes have been developed in the data stream community during 2001–2004. In contrast, most of the algorithms with for-all guarantees have been discovered after 2004, during the process of unifying the ideas of compressive sensing and data stream algorithms. We present the algorithms in the same chronological order.

Almost all schemes described here offer sketch length bounds of Oðklog nÞ or less, which matches or is close to the lower bound of Qðk logðn=kÞÞ shown in [4]. They are supported by efficient algorithms, with running times ranging from polynomial in n to near-linear in n. They offer a variety of approximation guarantees, starting from a"plain vanilla" guarantee of (1) with p 1/4 q 1/4 1 (the l1=l1-guarantee) to more complex (but often stronger) ones. The exact sketch length bounds, approximation guarantees, and algorithm running times are stated in Theorems 1–10.

Due to lack of space, we focus on describing only those algorithms that achieve the best known bounds and solve the sparse recovery problem formulated earlier in this section. See [19] and [20] for some of the earlier work on closely related problems, such as recovering good piece-wise constant approximations from a sketch of a signal.

II. ALGORITHMS WITH

FOR-EACH GUARANTEES

In this section, we describe algorithms that provide for-each guarantees. The algorithms were discovered and described in the context of data stream computing. The descriptions provided here are sometimes simpler than the original versions, since we ignore various issues specific to data streams (such as how to generate the random matrix A using few random bits, how to update the sketch under incremental changes to x, etc.).

A. Count-Min and Count-Median

The count-min and count-median algorithms [9] utilize

sparse random matrices where each entry is either 0 or 1.

 

Both algorithms use the same distribution of matrices, and differ only in the details of the recovery algorithms.

Each matrix A is generated in the following way. Let w be a parameter, and let h be any function from the set H of all functions h : f1; ... ; ng ! f1; ... ; wg. Each such function defines a w  n 0–1 matrix AðhÞ, such that ðAðhÞÞj;i is equal to 1 if j 1/4 hðiÞ, and is equal to 0 otherwise. Note that each column has exactly one 1.

To create the matrix A, we choose d functions h1; ... ; hd independently and uniformly at random from H. Then, we define A to be a vertical concatenation of matrices Aðh1Þ; ... ; AðhdÞ. Note that the number of rows in the matrix A is equal to m 1/4 wd.

For intuition about the construction, observe that, for any signal x, and j 1/4 1;... ; w, l 1/4 1; ... ; d, we have

ðAxÞðl1Þwþj 1/4 ðAðhlÞxÞj1/4 ~ xi:

i:hlðiÞ1/4j

That is, the coordinate of the sketch corresponding to the function hl and value j is simply the sum of all values xi such that i is mapped to j by hl. For a fixed value of i, the sums Et:hlðtÞ1/4hlðiÞ xt contain approximations of xi, contam¬inated by other coordinates mapped together with i. As a result, "aggregating" those sums over different hl provides an approximation of xi. Different aggregation methods will lead to different algorithms.

Count-Min: The count-min algorithm [9] (see also [18]) works under the assumption that x  0. In this case, computing the approximation x from Ax is particularly simple: we define

xi 1/4 minðAðhlÞxÞhlðiÞ1/4 min

l l

i0:hlði0Þ1/4hlðiÞ

xi0:

The guarantees for the estimator x can be derived as follows. First, observe that xi  xi , since the entries xi0 contaminating the estimation of xi can only increase the value of the estimator xi . Thus, the estimator ðAðhlÞxÞhlðiÞ with the minimum value provides the smallest approxi¬mation error. Moreover, for any coordinate xi and function index l, we have

E ðAðhlÞxÞhlðiÞxi 1/4

[ lJ E

i61/4i0 Pr1/2hlðiÞ 1/4 hlði0Þxi0

 1

w

kxk1:

By Markov inequality

~ 1

Pr ðAðhlÞxÞhlðiÞxi  w 2 kxk1  1=2

Vol. 98, No. 6, June 2010 | PROCEEDINGS OF THE IEEE 939

 

Gilbert and Indyk: Sparse Recovery Using Sparse Matrices

 

and, therefore

I J

Pr xi xi  2 kxk1  1=2d:

w

For d 1/4 Clog n, we have that the above guarantee holds for all i 1/4 1; ... ; n with probability 1  n=2d 1/4 1  1=nC1. Thus, with the same probability, we have

2

kx  xk1  kxk1:

w

The disadvantage of the above guarantee is that the error is a function of the norm of the whole vector x, not its tail. However, the probability that any of the entries in the head of x contaminate an estimator of a specific xi is at most k=w. Thus, a slightly more refined analysis5 shows that, for w 1/4 4=  k,  2 ð0; 1Þ, we have

[

Pr x  xi  =k  Errk]  1=2d:

1

For d 1/4 Clog n, this implies

kx  xk1  =k  Errk1

with probability 1  n=2d 1/4 1  1=nC1.

Count-Median: The count-min algorithm can be extend-ed to work for general signals [9]; the extension is often referred to as the count-median algorithm. The main issue to take care of is that, for general vectors x, the inequality xi  xi no longer holds, since the entries contaminating the estimator might be negative. As a result, we cannot aggregate using min. Instead, we replace the estimator x by

( )

x med i1/4 medianl ðAðhlÞxÞhlðiÞ:

By using the Chernoff bound, we show that, with high probability, the majority of the estimators ðAðhlÞxÞhlðiÞ (and therefore their median) have small error. Specifically, we can show that for any constant C0 > 0, there exists C such that if we set d 1/4 Clog n, then

~ ~

~ xmed  x~1 =k  Errk1

with probability 1  1=nC0.

5The argument is essentially a simplified version of the argument used in [6]. See [10] or [24, Lecture 4] for the proof.

940 PROCEEDINGS OF THE IEEE | Vol. 98, No. 6, June 2010

 

Theorem 1: There exists a distribution over m  n matrices A, m 1/4 Oðk=  log nÞ, such that for any signal x, given Ax, we can recover x^ 1/4 xmed such that

k^x  xk1  =k  Errk1

with high probability. The column sparsity of A is Oðlog nÞ, and the time needed to recover x^ from Ax is Oðn log nÞ.

We conclude by observing that the approximation guarantee in the above theorem implies a weaker but perhaps more intuitive guarantee about the l1 approxima¬tion error. Consider the vector x^ consisting of the k largest (in magnitude) elements of xmed. Then, we have

kx  ^xk1  ð1 þ 3ÞErrk1:

To show this, let S be the set of the k largest in magnitude coordinates of x, and let S^ be the support of ^x. Note that k^xSk1  k^x^Sk1. We have

kx  ^xk1  kxk1  kx^Sk1 þ kx^S  ^x^Sk1

 kxk1  k^x^Sk1 þ 2Errk1

 kxk1  k^xSk1 þ 2Errk1

 kxk1  kxSk1 þ 3Errk1

 ð1 þ 3ÞErrk1:

For more detailed descriptions of the algorithms, see [6], [9], and [18].

B. Count-Sketch

The next6 algorithm, called count-sketch [6], provides error guarantees that are a function of Errk2 as opposed to Errk1. This is accomplished by using a distribution over matrices A very similar to those used by count-min, with one difference: each nonzero entry is chosen indepen¬dently and uniformly at random from f1; 1g (instead just being equal to 1). Formally, let ri;l be independent random variables with values chosen uniformly at random from f1; 1g, and let the functions h1; ... ; hd be defined as in the previous section. Then, the matrix A is a vertical concatenation of matrices Aðh1Þ; ... ; AðhdÞ, where ðAðhlÞÞj;i is equal to ri;l if j 1/4 hlðiÞ, and is equal to 0 otherwise. To estimate the coordinate xi, one then uses the median estimator

xmed0 1/4 medianl ri;lðAðhlÞxÞhlðiÞ:

6Chronologically, the count-sketch algorithm has been invented before count-min. It is easier, however, to describe the ideas in the reverse order.

 

Gilbert and Indyk: Sparse Recovery Using Sparse Matrices

 

The analysis of count-sketch relies on the observation that

[( ) )2

Pr ri;lAðhl Þx hlðiÞ xi  C=w Errk  1=4

2

for some absolute constant C > 0. The final guarantee is captured by the following theorem.

Theorem 2: There exists a distribution over m  n matrices A, m 1/4 Oðk= log nÞ, such that for any signal x, given Ax, we can recover x^ such that

( )2

k^x  xk2 1  =k  Errk 2

with high probability. The column sparsity of A is Oðlog nÞ, and the time needed to recover x^ from Ax is Oðnlog nÞ.

As before, the approximation guarantee in the theorem implies a weaker but more intuitive guarantee, this time about the l2 approximation error. Consider the vector x^ consisting of the k largest (in magnitude) elements of x

med0. Then, we have [11]

kx ^xk22  ð1 þ 9p---Þ Errk

( )2:

2

The proof proceeds as follows. Let E 1/4 Errk2. Let S be the set of k largest (in magnitude) coordinates of x, and let S^ be the support of ^x. Moreover, for any set P, let P denote the complement of P. We have

~ ~

kx  ^xk22  ðx  ^xÞ^S

~ ~22þkxS^Sk2 2

2 þ xðS[^SÞ

2: (2)

The first term is bounded by k=k  E2 1/4 E2. To bound the second term, we proceed as follows. Consider any i 2 S  S^and j 2 S^  S. We have

 -------  ------ 

jxij  jxjj  j^xij  j^xjj þ 2 =kE  2 =kE: (3)

Let a 1/4 maxi2S^S jxij and b 1/4 minj2^SS jxjj. From (3), we  ------ 

have a  b þ 2 =kE. Thus

 ------ 

kxS^Sk2 2  a2jS  ^Sj  ðb þ 2 =kEÞ2jS  ^Sj:

 

Since kx^SSk2 2  b2j^S  Sj 1/4 b2jS  ^Sj, we continue

( 1------------- ------- J2

kxS^Sk22  kx^SSk2 jS  ^Sj þ 2 =kE jS  ^Sj  kx^SSk2 þ 2 -- 

( p E)2

 kx^SSk2 2 þ 4kx^SSk2 p E þ 4E2

-- 

 kx^SSk2 2 þ 4 -- 

p E2 þ 4E2

kx^SSk22 þ 8p --- E2:

Plugging into (2), we get

kx  ^xk2 2  E2 þ kx^SSk22 þ 8 -- 

p E2 þ xðS[^SÞ

2

2

 9 p E2 þ kxSk2

-- 

2

1/4 ð1 þ 9 p ÞE2:

-- 

C. Sublinear Algorithms

The above algorithms all run in time at least linear in the signal size as they entail estimating a value for each coordinate in the signal, even those that are insignificant. If our goal is to just report k nonzero terms of k-sparse approximation, then it is sufficient to find (or approxi¬mate) the top k values only to achieve similar error guarantees. Sublinear algorithms aim to do just that and to do so in time that scales polynomially with the number of terms k desired and logarithmically with the length of the input signal.

We start with the simplest example of a sublinear algorithm and its associated binary measurement matrix to find the unique nonzero entry in a signal of length n and sparsity 1. Let B be the binary matrix with ith column given by the binary representation of i, beginning with the first column i 1/4 0. We refer to this matrix as a bit-tester matrix.7 We add a row of 1’s to the bit-tester matrix (to estimate the signal value) and refer to this matrix as B1. It has logðnÞ þ 1 rows and n columns and from the measurements B1x of a vector x with a single large entry, we can determine both the value of the entry and its position in time logðnÞ þ 1. The measurements are simply the position in binary plus an estimate of the signal value and the recovery algorithm is trivial. It also can be seen that a similar approach applies even if the signal x is not exactly 1-sparse, but contains some "small" amount of noise.

For general signals, the approach is to "augment" the algorithms and measurement matrix constructions from previous sections with the matrix B1. Recall that those algorithms used simple hash functions which map signal

7Readers familiar with coding theory might recognize B as the parity-check matrix of the Hamming code.

Vol. 98, No. 6, June 2010 | PROCEEDINGS OF THE IEEE 941

 

Gilbert and Indyk: Sparse Recovery Using Sparse Matrices

 

coordinates xi to rows j of the measurements. Implicit in the correctness proofs was the ability of those hash functions to isolate a few significant signal values from one another. More precisely, if h is chosen uniformly at random from a prespecified family H of hash functions, then

Pr [ h(i) = h(i0) ] = O(1) 

w

for some w; that is, the probability that positions i and i0 are hashed into the same measurement is low. Using arguments similar to those above, we can show that if there are only k large (or nonzero) entries that are hashed into more than k measurements, then with high probabil¬ity, a large fraction of the significant entries are hashed into separate measurements. We can view this process as a random masking of the original signal, leaving a signal with only one significant entry, to which we can apply the bit-tester matrix. More precisely, each row of our final matrix M is the pointwise (Hadamard) product between a

row in A and a row in B1. We say that M is the row tensor



product of B1 and A, M = B1 r A. Note that M has

approximately klog(n) rows.

Once we have a good estimate of a large fraction of the significant entries, we can subtract their contribution from the original measurements (exploiting the linearity of the measurement process algorithmically, in addition to its role in the application). We then repeat the process, using "fresh" measurements.

By using the above techniques, we obtain the following result [21].

Theorem 3: There exists a distribution over m x n matrices A, m = O(klog n), such that for any signal x, given Ax, we can recover x^ such that JJ^x - xJJ1 < CErrk1 with high probability. The column sparsity of A is O(logc n) for some constant c, and the time needed to recover x^ from Ax is polynomial in k and log n.

III. ALGORITHMS WITH FOR-ALL GUARANTEES

In this section, we describe algorithms that provide for-all guarantees. The algorithms have been discovered during the process of unifying the ideas of compressive sensing with those from data stream algorithms. The key part of that process has been to identify concrete properties that 1) hold for a random sparse matrix with a nonzero probability and 2) are sufficient to support efficient and accurate recovery algorithms.

One such property is based on the notion of graph expansion [2], [44]. Consider a bipartite graph G = G(A) between two node sets U and V, with IUI = n and IVI = m, such that an edge (i; j) belongs to G if and only if Aj;i = 1. Informally, such a graph is an expander, if each small

942 PROCEEDINGS OF THE IEEE | Vol. 98, No. 6, June 2010

 

enough set of the nodes in U has many neighbors in V (the formal definition is provided below).

The notion of expansion has been known to be useful for some related problems, such as constructing low-density parity-check codes. In fact, iterative decoding algorithms for such codes have been used, e.g., in [25], [27], and [44], to design sparse recovery algorithms. However, those algorithms were designed and proven to work only for the case where the signal x is either exactly k-sparse or "almost" k-sparse. In contrast, the algorithms we present here work for arbitrary input signals x.

Formally, we define unbalanced expander graphs as follows. Consider a bipartite graph G = (U; V; E), where E C U x V is the set of edges. We refer to U as the "left" part, and refer to V as the "right" part of the graph. A vertex belonging to the left (right) part is called a left (right) vertex. In our constructions, the left part will correspond to the set {1; 2; ... ; n} of coordinate indexes of vector x, and the right part will correspond to the set of row indexes of the measurement matrix. A bipartite graph is called left-d-regular if every vertex in the left part has exactly d neighbors in the right part.

Definition: A bipartite, left-d-regular graph G = (U; V; E) is an (s; d; )-expander if any set S C U of at most s left vertices has at least (1 - )dISI neighbors.

The algorithms described in this section use adjacency matrices A of the expanders graphs G: we simply set Aj;i = 1 if and only if (i; j) E E. Note that the resulting matrices are sparse, with exactly d ones per column.

What are the achievable expansion parameters? Since expander graphs are meaningful only when IVI<dIUI, some vertices must share neighbors, and hence the parameter  cannot be smaller than 1=d. Using the probabilistic method one can show that there exist (s; d; )-expanders with d = O(log(n=s)=) and m = IVI = O(slog(n=s)=2). Since our constructions require s = O(k) and  strictly bounded away from zero, the resulting matrices will have O(klog(n=k)) rows.

For many applications, one often needs an explicit expander, i.e., an expander for which we can efficiently compute the neighbor set of a given left vertex. No explicit constructions with the aforementioned parameters are known. However, it is known [23] how to explicitly construct expanders with left degree d = O((log IUI)(log s)=)1+1= and right set size (d2s1+), for any fixed  > 0. For simplicity, in the remainder of this paper, we will assume expanders with the optimal parameters.

Unlike in the for-each case,8 the algorithms in this section are known to be resilient to the measurement noise. That is, we could assume that we are given a noisy sketch vector b = Ax + , where  is the "measurement noise" vector. In that case, the error bounds in the approximation

8It should be noted that, although the for-each algorithms have not been typically analyzed for the case of noisy sketches, the algorithm themselves could very well be quite resilient to various forms of noise.

 

Gilbert and Indyk: Sparse Recovery Using Sparse Matrices

 

guarantees would have an additional term depending on  1/4 kk1=d. However, for the sake of consistency, we will focus the description on the noise-free case where b 1/4 Ax. The reader is referred to the original papers for the bounds for the noise-resilient variants of the algorithms.

A. RIP(1) and ‘1-Minimization

In this section, we give an overview of the "geometric" approach to sparse recovery using sparse matrices, introduced in [2]. The approach uses the ‘1-minimization algorithm that has been earlier shown to work for random dense matrices [12], [16]. In the noiseless case b 1/4 Ax, the algorithm proceeds by finding x^ such that A^x 1/4 b and k^xk1 is minimized.

To understand when the above algorithm performs an accurate recovery, we need the following generalized definition of the restricted isometry property.

Definition 2: An m  n matrix A is said to satisfy RIPðp; k; Þ if, for any k-sparse vector x, we have

kxkpð1  Þ  kAxkp  kxkp:

For the case of p 1/4 2, the notion was introduced9 in [12], which also showed that if a matrix A satisfies this property, then the ‘1-minimization procedure produces an accurate solution. Since then there has been a tremendous amount of study of the properties and construction of RIPð2; k; Þ [or RIP(2), for short] matrices. Unfortunately, sparse matrices cannot satisfy the RIP(2) property, unless their number of rows is "large" [8]. In particular, sparse 0–1 matrices must have at least Qðk2Þ rows.

However, it was shown [2] that such matrices can satisfy RIPðpÞ for p equal (or very close) to 1. In particular, the adjacency matrices of expander graphs do have this property.10 By earlier arguments, such matrices have Oðklogðn=kÞÞ rows, which translates into Oðklogðn=kÞÞ sketch length bound.

Lemma 4: Consider any m  n matrix A that is the adjacency matrix of an ðk; d; Þ-unbalanced expander G 1/4 ðU; V; EÞ. Then, the scaled matrix A=d satisfies the RIPð1; k; Þ property for  1/4 2.

Proof: Let x 2 Rn be a k-sparse vector. Without loss of generality, we assume that the coordinates of x are ordered such that jx1j  ...  jxnj. We order the edges et 1/4 ðit; jtÞ, t 1/4 1; ... ; dn of G in a lexicographic manner. It is helpful to imagine that the edges e1; e2;... of E are being

9The original paper [12] employed a slightly different notation using "double sided error," i.e., requiring that kxk2ð1  0Þ  kAxk2  kxk2ð1 þ 0Þ. The two definitions can be seen to be equivalent, by scaling A and setting ð1 þ Þ 1/4 ð1þ0Þ=ð1  0Þ.

10In fact, for some range of parameters, the opposite holds, i.e., 0–1 matrices that satisfy RIP(1) are adjacency matrices of expander graphs. See [2] and [8] for more details.

 

added to the (initially empty) graph. An edge et 1/4 ðit; jtÞ causes a collision if there exists an earlier edge es 1/4 ðis; jsÞ, s < t, such that jt 1/4 js. We define E0 to be the set of edges which do not cause collisions, and E00 1/4 E  E0.

Claim 5: We have

jxij  dkxk1:

ði;jÞ2E00

Proof: For each t 1/4 1;...; dn, we use an indicator variable rt 2 f0; 1g, such that rt 1/4 1 iff et 2 E00. Define a vector z 2 Rdn such that zt 1/4 jxitj. Observe that

 jxij 1/4 rt jxitj 1/4 r  z

ði;jÞ2E00 et1/4ðit ;jtÞ2E

To upper bound the latter quantity, observe that the vectors satisfy the following constraints:

• the vector z is nonnegative;

• the coordinates of z are monotonically nonincreas-ing, and at most kd of them are nonzero;

• for each prefix set Pi 1/4 f1; ... ; dig, i  k, we have krjPik1  diVthis follows from the expansion properties of the graph G;

• rjP1 1/4 0, since the graph is simple.

It follows that for any r; z satisfying the above

constraints, we have r  z  kzk1. Since kzk1 1/4 dkxk1,

the lemma follows. 

Since

kAxk1  jxit j  jxitj

et1/4ðit ;jtÞ2E0 et1/4ðit;jtÞ2E00

then Claim 5 immediately implies that kAxk1  dkxk1ð1  2Þ. Since for any x we have kAxk1  dkxk1, it

follows that A=d satisfies the RIPð1; k; 2Þ property. 

We now need to show that the RIP(1) property of the matrix A is sufficient to guarantee that the ‘1-minimization works. First, we show that any vector from the kernel of an adjacency matrix A of an expander graph (i.e., such that Ax 1/4 0) is "smooth," i.e., the ‘1-norm of the vector cannot be concentrated on a small subset of its coordinates. An analogous result for RIP(2) matrices and with respect to the ‘2-norm has been used before to show guarantees for LP-based recovery procedures.

Lemma 6: Consider any y 2 Rn such that Ay 1/4 0, and let S be any set of k coordinates of y. Then, we have

kySk1  ðÞkyk1

where ðÞ 1/4 ð2Þ=ð1  2Þ.

Vol. 98, No. 6, June 2010 | PROCEEDINGS OF THE IEEE 943

 

Gilbert and Indyk: Sparse Recovery Using Sparse Matrices

 

The proof proceeds by showing that any vector y whose Q1-norm is concentrated on a small set S of coordinates cannot satisfy Ay = 0. This is because [by the RIP(1) property] the Q1-norm of the vector A(yS) is "large," and (from the expansion property of the underlying graph) the contribution of the coordinates in the complement of S is not sufficient to reduce Ay to 0. See [2] for the formal proof.

The "smooth kernel" property is then used, as in prior work, to provide recovery guarantees for the e1-minimization. This is achieved by the following lemma, by setting u = x and v = ^x.

Lemma 7: Consider any two vectors u, v, such that for y = v - u, we have Ay = 0, and llvll1 < llull1. Let S be the set of k largest (in magnitude) coefficients of u, then

llv - ull1 < 2/(1- 2a(e)) • llu - uSll1.

The following theorem summarizes the discussion.

Theorem 8: There exists an m x n (expander) matrix A, m = O(k log(n/k)/e2), such that for any signal x, given Ax, we can recover x^ such that

llx - ^xll1 < c(c)Errk1

where c(c) -*2 as c -* 0. The column sparsity of A is O(log(n)/e2), and the recovery involves solving a linear program with O(n) variables and O(m + n) constraints.

This concludes the overview of the results of [2]. Further studies of e1-minimization algorithms for sparse matrices have been done in [43] and [29], where the authors obtained tight estimates for the number of measurements needed to recover signals of given sparsity. The papers consider somewhat different setups: in [43], one allows arbitrary sparse signals x and measurements contaminated by random Gaussian noise; in [29], the authors consider sparse nonnegative signals.

B. EMP, SMP, and Other Near-Linear

Time Algorithms

In this section, we describe a family of iterative algorithms for performing sparse recovery. Their key feature is that they enable performing sparse recovery in near-linear time while still using O(klog(n/k)) measure¬ments. The algorithms do not use linear programming; instead, they exploit various forms of voting mechanisms to converge to a solution. The specific algorithms covered are: expander matching pursuit (EMP) [26] and sparse matching pursuit (SMP) [5].11

11There is a very recent variant of SMP called sequential sparse matching pursuit [3]. We do not cover it in this survey due to lack of space.

944 PROCEEDINGS OF THE IEEE | Vol. 98, No. 6, June 2010

 

To describe the algorithms we need some notation. For a set S of nodes of a graph G, the ordered set of its neighbors in G is denoted by PG(S). The subscript G will be omitted when it is clear from the context, and we write P(u) as a shorthand for P({u}).

Both EMP and SMP proceed in a sequence of steps, where each step is similar to the median estimation process of the count-median algorithm. A minor technical difference is that the algorithms are constructed for general sparse matrices A, as opposed to block-structured ones used by count-median. Therefore, for a given sketch vector b = Ax, the median estimation vector Emed(b) is defined as

Emed(b)i = medianjEr(i)bj.

That is, each vertex i selects the entries bj where j is a neighbor of i in G, and then computes the median of those entries. One can observe that for the matrices used by count-median, the new and old estimators are identical. The basic intuitions behind the choice of the estimator transfer as well.

There is, however, one important difference: unlike in the for-each setup, here we cannot guarantee that each coordinate Emed(b)i differs from xi by only a small term. In fact, due to the deterministic nature of the process, it might be possible that, for some coordinate i, all sketch coordinates bj, j E P(i), could be highly "contaminated" by other entries of x. Thus, the algorithms do not provide guarantees for the lo. error of the recovered approxima¬tion. However, it is nevertheless possible to directly give guarantees for the l1 approximation error.

1) EMP: The first algorithm that achieved the O(klog(n/k)) sketch length bound and recovery time near-linear in n was the EMP. The algorithm and its analysis are somewhat complicated, so instead of a detailed description we provide only an overview.

EMP consists of two phases. In the first phase, the algorithm identifies a set I of coordinates of x that 1) contains "most" of the k largest (in magnitude) coeffi-cients of x and 2) for all nodes i E6 I the neighborhood sets P(i) and P(I) have "small" intersection. The first constraint ensures that we can set the coordinates ^xi of the approximation to zero for all i E6 I. The second constraint ensures that the values of sketch coordinates br(I) are not too contaminated by entries xi for i E6 I. Together, this implies that we can focus on decoding ^xI from br(I). This is accomplished during the second phase, which proceeds in a sequence of iterations. In each iteration, the algorithm identifies coordinates i E I such that most of elements of P(i) do not have any other neighbors in P(I). The algorithm then estimates the values ^xi of such coordinates (using the median estimator), eliminates them from the I, and subtracts their

 

Gilbert and Indyk: Sparse Recovery Using Sparse Matrices

 

contribution to the sketch. The process is continued until the set I becomes empty.

Since each coordinate of the approximation is estimat¬ed only once, and is never revised again, the EMP algorithm is very efficient: it runs in time proportional to the number of edges in the graph G, which is O(nlog(n/k)). The recovered vector x^ provides an approximation in the 21-norm, i.e., we have the following.

Theorem 9: There exists an m x n (expander) matrix A, m = O(k log(n/k)/a2), such that for any signal x, given Ax, we can recover x^ such that

IIx - ^xII1 < (1 + a)Errk1.

The column sparsity of A is O(log(n)/a2), and the recovery algorithm (EMP) has O(nlog(n/k)/a2) running time.

Although EMP offers excellent asymptotic guarantees, its empirical performance is not so great. Specifically, the number of measurements required by the algorithm to achieve correct recovery is suboptimal. For example, our recovery experiments on random signed k-sparse signals of length n, for k = 50 and n = 20 000, show that one typically needs at least 5000 measurements to recover the signal correctly using the EMP algorithm. In comparison, the linear-programming-based recovery algorithm for sparse matrices described earlier requires only about 450 measurements to perform the same task.12

2) SMP: The SMP borrows some of the ideas present in EMP, but it has been also influenced by the recent iterative algorithms for sparse recovery using dense matrices, such as [35]. The running time of the new algorithm is slightly higher (by a logarithmic factor) than of EMP. However, empirically, the algorithm performs successful recovery from a significantly smaller number of measurements. In particular, for the instances described above, SMP typically needs about 2000 measurements. The asymptotic bound on the number of required measurements is still O(klog(n/k)).

The recovery algorithm is iterative, in the spirit of matching pursuit [40]. In each iteration, the algorithm estimates the difference between the current approxima¬tion ^xj and the signal x from the sketch A^xj - b. The estimation, denoted by u*, is obtained by using the median estimator as in EMP. The approximation ^xj is updated by u, and the process is repeated.

Let Hl[y] be a "thresholding operator," which zeros out all but the l largest in magnitude coefficients of the argument y. Also, let C > 0 be some constant. The details of the algorithm, together with remarks about the properties used in the analysis, are depicted in Fig. 1.

The remarks rely on the following trick, borrowed from [35]: we can decompose the input signal x into the "head"

12For both algorithms, we used randomly generated 0–1 matrices with column sparsity equal to 20.

 

 

Fig. 1. Sparse matching pursuit algorithm: pseudocode and remarks on the analysis.

x' (containing the k most significant components of x) and the "tail" x - x'. Then, we can interpret the "sketch of the tail" term A(x - x') as measurement noise. That is, we can assume that the sketch b is equal to Ax' + µ', where µ' = A(x - x') and x' is k-sparse. Note that the RIP(1) property of A implies that IIA(x - x')II1 < dIIx - x'II1 = dErrk1. We define T' = IIµ'II1/d < Errk1.

From the remarks in the algorithm description, we conclude that for any j = 1, 2, ... , T, we have

II^xj - x'II1 < IIx'II1/2j + O(T').

Thus, setting the number of iterations to T = log(IIx'II1/T') guarantees that

II^xT - x'II1 = O(T') = O Errk

( ).

1

The following theorem summarizes the discussion.

Theorem 10: There exists an m x n (expander) matrix A, m = O(klog(n/k)), such that for any signal x, given Ax, we can recover x^ such that

IIx - ^xII1 < c Errk1

for an absolute constant c > 0. The column sparsity of A is O(logn), and the recovery algorithm (SMP) has O(nlog(n/k)T) running time, for T defined as above.

3) Connections to Message-Passing Algorithms: The SMP algorithm described above, as well as the aforementioned algorithms from [25], [27], and [44], can be interpreted in

Vol. 98, No. 6, June 2010 | PROCEEDINGS OF THE IEEE 945

 

Gilbert and Indyk: Sparse Recovery Using Sparse Matrices

 

a general framework of message-passing algorithms. Such algorithms structure their operations based on the bipartite graph G underlying the matrix A. Specifically, each node of the graph can be viewed as a separate processing unit, and the algorithm proceeds by the units sending messages to each other along the edges of the graph. Message-passing algorithms have numerous advan¬tages over the "centralized" ones: their computational complexity is low (if the underlying graph is sparse); they also can be easily implemented in a parallel or distributed manner.

There have been several papers on message-passing algorithms for sparse recovery problems using sparse random matrices. In [37] and [39], the authors introduced the belief propagation approach to compressive sensing, and applied it to the recovery of random signals, modeled by a two-state mixture of Gaussians. In a more recent paper [1], the authors used belief propagation on signals modeled as Gaussian-scale mixtures to obtain algorithms with an excellent empirical performance.

Message passing framework has been also used to design randomized algorithms that work in the worst case. In particular, the paper [32] introduced and analyzed such algorithms that work for arbitrary k-sparse signals. That algorithm can be viewed as an iterative generalization of the count-min algorithms described in earlier sections.

C. HHS and Sublinear Algorithms

As in Section II, there are versions of the above algorithms with sublinear running times. The main example is heavy hitters on steroids (HHS) [22]. The output of the

V

HHS algorithm is x^ where 11x - ^x112 < C(Err2 + 1/ kErr1) and its running time is k2(log n)O(1).It retains the same overall architecture as the iterative algorithms: within each 

 

step, it isolates significant entries by hashing, estimates their values, and then updates the measurements accordingly. It shares a "voting" procedure for determining significant signal entries with the EMP and SMP algorithms; however, these votes are derived from the bit tests rather than from the signal estimates directly. HHS differs from the simple sublinear algorithm we sketched in Section II in three major parts. First, in order to obtain a strong guarantee for all signals, we must hash k significant entries into O(k) measurements repeatedly, for O(logn) repetitions. The adjacency matrix of a (s, d, e) expander with s = O(k) is a way to achieve this. Second, because we use a simple bit tester B1 to identify the significant entries, we must ensure that it is applied to a signal that is sufficiently filtered; the contribution of the insignificant entries must be small enough not to pollute our estimates of the significant entry (recall that because the algorithm is iterative, estimation errors at one stage can accumulate at further iterations). Furthermore, we must carefully balance the ‘1 and ‘2 errors. To this end, we employ a second hash matrix that reduces the noise in each measurement after the first hash. In each iteration j, we keep a list of signal positions for which we have



at least k/j log klog(n/j)log(n) votes. Third, we use a separate matrix to estimate the values of the identified signal positions with the desired mixed norm error guarantee. Finally, in each iteration, we prune the list of signal positions to retain the top O(k) positions. h

Acknowledgment

The authors would like to thank J. Nelson, G. Cormode, and the anonymous reviewers for very helpful and insightful comments.

 

 

REFERENCES

[1] M. Akcakaya, J. Park, and V. Tarokh. (2009). Compressive sensing using low density frames. [Online]. Available: http://arxiv.org/ abs/0903.0650

[2] R. Berinde, A. Gilbert, P. Indyk, H. Karloff, and M. Strauss, "Combining geometry and combinatorics: A unified approach to sparse signal recovery," in Proc. 46th Annu. Allerton Conf. Commun. Control Comput., 2008, pp. 798–805.

[3] R. Berinde and P. Indyk, "Sequential sparse matching pursuit," in Proc. 47th Annu. Allerton Conf. Commun. Control Comput., 2009, pp. 36–43.

[4] K. Do Ba, P. Indyk, E. Price, and D. Woodruff, "Lower bounds for sparse recovery," in Proc. 19th Annu. ACM-SIAM Symp. Discrete Algorithms, 2010.

[5] R. Berinde, P. Indyk, and M. Ruzic, "Practical near-optimal sparse recovery in the l1 norm," in Proc. 46th Annu. Allerton Conf. Commun. Control Comput., 2008, pp. 198–205.

[6] M. Charikar, K. Chen, and M. Farach-Colton, "Finding frequent items in data streams," in 

 

Proc. Int. Coll. Autom. Lang. Programm., pp. 693–703, 2002.

[7] G. Cormode and M. Hadjieleftheriou, "Finding the frequent items in streams of data," Commun. ACM, vol. 52, pp. 97–105, 2009.

[8] V. Chandar, "A negative result concerning explicit matrices with the restricted isometry property,”2008, preprint.

[9] G. Cormode and S. Muthukrishnan, "Improved data stream summaries: The count-min sketch and its applications,”J. Algorithms, vol. 55, no. 1, pp. 58–75, 2005.

[10] G. Cormode and S. Muthukrishnan, "Summarizing and mining skewed data streams," in Proc. SIAM Int. Data Mining Conf., 2005.

[11] G. Cormode and S. Muthukrishnan, "Combinatorial algorithms for compressed sensing," in Proc. 40th Annu. Conf. Inf. Sci. Syst., Princeton, NJ, Mar. 2006, pp. 198–201.

[12] E. Cande`s, J. Romberg, and T. Tao, "Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information," IEEE Trans Inf. 

 

Theory, vol. 52, no. 2, pp. 489–509, Feb. 2006.

[13] M. Duarte, M. Davenport, D. Takhar, J. Laska, T. Sun, K. Kelly, and R. Baraniuk, "Single-pixel imaging via compressive sampling," IEEE Signal Process. Mag., vol. 25, no. 2, pp. 83–91, Mar. 2008.

[14] R. DeVore, "Deterministic constructions of compressed sensing matrices,”2007, preprint.

[15] D.-Z. Du and F. K. Hwang, Combinatorial Group Testing and Its Applications. Singapore: World Scientific, 1993.

[16] D. L. Donoho, "Compressed sensing," IEEE

Trans. Inf. Theory, vol. 52, no. 4,

pp. 1289–1306, Apr. 2006.

[17] Y. Erlich, N. Shental, A. Amir, and O. Zuk, "Compressed sensing approach for high throughput carrier screen," in Proc. 47th Annu. Allerton Conf. Commun. Control Comput., 2009, pp. 539–544.

[18] C. Estan and G. Varghese, "New directions in traffic measurement and accounting: Focusing on the elephants, ignoring the mice," ACM Trans. Comput. Syst., vol. 21, no. 3, pp. 270–313, Aug. 2003.

 

946 PROCEEDINGS OF THE IEEE | Vol. 98, No. 6, June 2010

 

Gilbert and Indyk: Sparse Recovery Using Sparse Matrices

 

[19] A. C. Gilbert, S. Guha, P. Indyk, Y. Kotidis, S. Muthukrishnan, and M. J. Strauss, "Fast, small-space algorithms for approximate histogram maintenance," in Proc. 34th Annu. ACM Symp. Theory Comput., 2002, pp. 389–398.

[20] A. C. Gilbert, Y. Kotidis, S Muthukrishnan, and M. Strauss, "One-pass wavelet decompositions of data streams," IEEE Trans. Knowl. Data Eng., vol. 15, no. 3, pp. 541–554, May/Jun. 2003.

[21] A. Gilbert, Y. Li, E. Porat, and M. Strauss. (2009). Approximate sparse recovery: Optimizing time and measurements. [Online]. Available: http://arxiv.org/abs/0912.0229

[22] A. C. Gilbert, M. J. Strauss, J. A. Tropp, and R. Vershynin, "One sketch for all: Fast algorithms for compressed sensing," in Proc. 39th Annu. ACM Symp. Theory Comput., 2007, pp. 237–246.

[23] V. Guruswami, C. Umans, and S. P. Vadhan, "Unbalanced expanders and randomness extractors from Parvaresh-Vardy codes," in Proc. IEEE Conf. Comput. Complex., 2007, pp. 96–108.

[24] P. Indyk. (2007). Sketching, streaming and sublinear-space algorithms. graduate course notes. [Online]. Available: http://stellar.mit. edu/S/course/6/fa07/6.895/

[25] P. Indyk, "Explicit constructions for compressed sensing of sparse signals," in Proc. 19th Annu. ACM-SIAM Symp. Discrete Algorithms, 2008, pp. 30–33.

[26] P. Indyk and M. Ruzic, "Near-optimal sparse recovery in the l1 norm," in Proc. 49th Annu. Symp. Found. Comput. Sci., 2008, pp. 199–207.

[27] S. Jafarpour, W. Xu, B. Hassibi, and A. R. Calderbank. (2008). Efficient and robust compressed sensing using high-quality expander graphs. [Online]. Available: http:// arxiv.org/PS_cache/arxiv/pdf/0806/0806. 3802v1.pdf

 

[28] R. Kainkaryam, A. Bruex, A. Gilbert, P. Woolf, and J. Schiefelbein, “Poolmc: Smart pooling of

MRNA samples in microarray experiments,”2010, Manuscript.

[29] M. A. Khajehnejad, A. G. Dimakis, W. Xu, and B. Hassibi. (2008). Sparse recovery of positive signals with minimal expansion. [Online]. Available: http://arxiv.org/abs/0902.4045

[30] B. Krishnamurthy, S. Sen, Y. Zhang, and Y. Chen, "Sketch-based change detection," in Proc. 3rd ACM SIGCOMM Conf. Internet Meas., 2003, pp. 234–247.

[31] J. Laska, S. Kirolos, Y. Massoud, R. Baraniuk, A. Gilbert, M. Iwen, and M. Strauss, "Random sampling for analog-to-information conversion of wideband signals," in Proc. IEEE Dallas

CAS Workshop Design Appl. Integr. Softw., 2006, pp. 119–122.

[32] Y. Lu, A. Montanari, B. Prabhakar,

S. Dharmapurikar, and A. Kabbani, "Counter braids: A novel counter architecture for per-flow measurement," in Proc. ACM SIGMETRICS Int. Conf. Meas. Model. Comput. Syst., 2008, pp. 121–132.

[33] S. Muthukrishnan. (2003). Data streams: Algorithms and applications, Invited Talk at the 14th Annu. ACM-IAM Symp. Discrete Algorithms. [Online]. Available: http://athos. rutgers.edu/~muthu/stream-1-1.ps

[34] S. Muthukrishnan, "Some algorithmic problems and results in compressed sensing," in Proc. Annu. Allerton Conf. Commun. Control Comput., 2006.

[35] D. Needell and J. A. Tropp, "CoSaMP: Iterative signal recovery from incomplete and inaccurate samples," Appl. Comput. Harmonic Anal., vol. 26, no. 3, pp. 301–321, 2009.

[36] N. Shental, A. Amir, and

O. Zuk. (2009). Rare-Allele detection using compressed se(que)nsing. [Online]. Available: arXiv:0909.0400

 

[37] S. Sarvotham, D. Baron, and R. G. Baraniuk, "Compressed sensing reconstruction via belief propagation," Electr. Comput. Eng. Dept., Rice Univ., Houston, TX, Tech. Rep. ECE-0601, 2006.

[38] S. Sarvotham, D. Baron, and R. G. Baraniuk, "SudocodesVFast measurement and reconstruction of sparse signals," in Proc. IEEE Int. Symp. Inf. Theory, 2006.

[39] S. Sarvotham, D. Baron, and

R. G. Baraniuk. (2008). Bayesian compressive

sensing via belief propagation. [Online].

Available: http://arxiv.org/abs/0812.4627

[40] J. A. Tropp and A. C. Gilbert, "Signal recovery from random measurements via orthogonal matching pursuit," IEEE Trans. Inf. Theory, vol. 53, no. 12, pp. 4655–4666, Dec. 2007.

[41] J. Tropp, M. Laska, M. Duarte, J. Romberg, and R. Baraniuk, "Beyond Nyquist: Efficient sampling of sparse bandlimited signals," IEEE Trans. Inf. Theory, vol. 56, no. 1, pp. 520–544, Jan. 2010.

[42] D. Takhar, J. Laska, M. B. Wakin,

M. F. Duarte, D. Baron, S. Sarvotham, K. Kelly, and R. G. Baraniuk, "A new compressive imaging camera architecture using optical-domain compression," in Proc. IS&T/SPIE Symp. Electron. Imaging, 2006, vol. 6065, DOI:10.1117/12.659602606509.

[43] W. Wang, M. J. Wainwright, and K. Ramchandran. (2008). Information-theoretic limits on sparse signal recovery: Dense versus sparse measurement matrices. [Online]. Available: http://arxiv.org/abs/0806.0604

[44] W. Xu and B. Hassibi, "Efficient compressive sensing with deterministic guarantees using expander graphs," in Proc. IEEE Inf. Theory Workshop, 2007, pp. 414–419.

 

 

ABOUT THE AUTHORS

Anna Gilbert received the S.B. degree from the University of Chicago, Chicago, IL and the Ph.D. degree from Princeton University, Princeton, NJ, in 1997, both in mathematics.

In 1997, she was a Postdoctoral Fellow at Yale University and AT&T Labs-Research. From 1998 to 2004, she was a member of the technical staff at AT&T Labs-Research, Florham Park, NJ. In 2004, she joined the Department of Mathematics, University of Michigan, Ann Arbor, where she is now an Associate Professor. Her research interests include analysis, probability, networking, and algorithms. She is especially interested in randomized algorithms with applications to harmonic analysis, signal and image processing, networking, and massive data sets.

Dr. Gilbert received a National Science Foundation (NSF) CAREER Award, a Sloan Research Fellowship, and a NAS Award for Initiatives in Research. 

 

Piotr Indyk received the M.S. degree from Uniwersytet Warszawski, Warsaw, Poland, in 1995 and the Ph.D. degree from Stanford University, Stanford, CA, in 2000, both in computer science.

Currently, he is an Associate Professor of Electrical Engineering and Computer Science at the Massachusetts Institute of Technology (MIT), Cambridge. His research interests include high-dimensional computa¬tional geometry, sketching and streaming algorithms, sparse recovery and compressive sensing.

Dr. Indyk received a National Science Foundation (NSF) CAREER Award, a Sloan Fellowship, and a Packard Fellowship.

 

Vol. 98, No. 6, June 2010 | PROCEEDINGS OF THE IEEE 947

 

DEVELOPMENTS IN COMPUTER TECHNOLOGY IN THE

UNIVERSITY OF THE PHILIPPINES: USERS' EXPECTATIONS AND NEEDS

Evangeline C. Capuli

Marine Science Institute, University of the Philippines

UPPO Box 1, Diliman, 1101 Quezon City, Philippines

ABSTRACT: The widespread use of computer technology has spawned the expectations of users for speedy, efficient, and useable information because they are no longer satisfied with the traditional flow of the current surplus of information materials. In the University of the Philippines system, this need for access to useable information has materialized from recent subscriptions of CD-ROMs and from linkage with the Internet through the Engineering and Science Education Project (ESEP) of the Department of Science and Technology (DO ST). However, before the potential benefits brought about by these innovations can be fully realized, a number of barriers have to be overcome, among which are the misguided perception of library managers on the role of the network as information providers and the lack of handling skills by concerned personnel.

Introduction

In response to a dynamic information environment, users are now aware of the availability of varied information and communication technology suited to their individual requirements. These technological advances, which include computerized databases, CD-ROMs, telecommunication networks, electronic publishing, fax machines and photocopiers, have a great impact on library services, most notably in the areas of document delivery and interlibrary loan.

Background: electronic document delivery is a hot topic

In libraries of academic institutions such as the University of the Philippines, users rely heavily on serial articles to such a degree that the current percentage of requests for interlibrary loan and delivery of journal articles is much higher than in any other type of library. In a study conducted in the mid- 80's, a great majority of requests for serial articles was met with the provision of photocopies of the requested papers (Walhart 1985, cited in Kinnucan 1993). Before the introduction of the photocopying machine, it could thus be surmised how helpless libraries were in the midst of a deluge of interlibrary loan requests for journal articles since most were not, and still are not up to this present day and age, inclined to lend entire journal issues, especially bound volumes.

Telecommunication networks and fax machines have similarly accelerated the provision of requested documents for libraries that use them (Medina 1992, cited in Kinnucan 1993). The

 

51

 

introduction of computerized databases has made it possible for information users to learn of information they might have otherwise overlooked and has made it easier for librarians as well to locate which library has the needed document.

The above-mentioned innovations have undoubtedly changed the way interlibrary loan is accomplished and have made document delivery a reality (Kinnucan 1993).

Current advances

Worldwide initiatives in the information field are conceived to facilitate easy access for researchers and practitioners to their required information. The introduction of the CD-ROM in the mid-80s has made possible the access and dissemination of a large amount of bibliographic, full text and other data. CD-ROM titles are now available in the market covering a wide range of scientific disciplines. Simultaneously, the development of new interfaces and their continued update aid in extending CD-ROM feature services to a wider spectrum of the information community. This has therefore caused a change in the management of electronic searching. Librarians who were, to a certain point, considered intermediaries of information now had to focus their attention on training inexperienced users to search electronically.

As CD-ROM's popularity grew, so did the numbers of users and their different searching needs. As a result, experienced users requested features that would allow them to navigate and find exactly the information they needed from the available databases while those not as familiar with the new technology wanted the search kept simple.

However, while acknowledging the definite advantages of computer technologies in which CD-ROMs operate, Keylard (1993) mentions a few caveats that may lead to the under-utilization of CD-ROMs:

a. the misconception of end-users regarding a particular computer technology such as the CD-ROM;

b. the assumption that users and intermediaries require no training;

c. lack of understanding of the technology involved;

d. lack of integration of said technologies in library services; and

e. lack of long-term provision.

These so-called 'pitfalls' can easily be avoided or overcome by information campaigns initiated by either intermediaries or unit heads, appropriate training and, an assurance of funding support. A sound implementation plan needs to be conceived prior to the employment of such technologies. The parent organization should also make a commitment to integrate such technologies within the various library services to ensure their maximum use.

 

52

 

Information and the Internet The UPIDOST Engineering and Science Education Project (ESEP)

Dubbed as the "mother of all networks", Internet is the world's largest computer network linking other networks in various countries. According to Steven Goldstein, Internet Program Director, more than 30,000 autonomous networks registered with the National Science Foundation Network (NSFNET) have linked to Internet in 1993 (Magno 1994).

The Philippines' link to Internet was finalized in December 1993 following an allocation of P1 2.4M from the Department of Science and Technology (DOST) (Hilotin 1994, Magno 1994). This amount, equivalent to roughly half a million dollars, was for the installation and financing of PhilNet, the node which would serve as the central computer station through which remote computers can access the Internet's resources.

Institutions or individuals from the academic, research and commercial sectors connect to PhilNet on a subscription basis. The connection rates depend on the classification of the user, type of connection and speed of lines. For ESEP, presently in place are the access nodes in the University of the Philippines (UP) in Diliman, Ateneo de Manila University (ADMU), De La Salle University (DLSU), University of Santo Tomas (UST), UP Manila, UP at Los Banos, Saint Louis University in Baguio City, University of San Carlos (USC) in Cebu City, Xavier University in Cagayan de Oro City, Mindanao State University (MSU) in lligan, and the DOST.

It is widely believed that such direct link to Internet will hasten the transfer of scientific discoveries to and from the Philippines, since it takes months for such data to be printed in a scientific journal, stored in a library, and disseminated to the scientific communities of developing countries. Outside of this network, however, very few librarians possess the adequate know-how on automated library systems, Internet, and other applications of information technologies. To ensure the creation of a truly nationwide network, training courses are being proposed to update the technological know-how of these librarians.

Status of library networks in the University of the Philippines

Library networks leave their greatest impact on the librarians and users alike. Librarians have been offered more challenging opportunities for better service as new technology replaced their usual manual routines, i.e., roles as negotiators, facilitators, educators, and information brokers (SMeart, 1982). These roles, nevertheless, require training and skills development for effective database handling and better appreciation of the technology, to ensure successful integration of the technology into the traditional library services being offered.

On the other hand, users are able to access remote information in so short a time. Users are assumed, however, to understand different communication media. But since network systems are

 

53

 

designed for self-help, users are faced with either of the following obstacles: non-readiness, lack of computer skills, and worse, alienation. Users would then have to seek assistance from either the information managers or intermediaries. Hands-on seminars regularly organized, e.g., each semester or year, may become a necessity for a growing number of users. User-friendly manuals and computer routines may be produced since these can be avenues of self-training. The latter two, for our case, remain to be realized in the near future.

Conclusion and Recommendations

With the coming of what is known as a global information infrastructure (Valentin 1995) one can foresee a scenario for the world of information that is "high-tech" and highly profitable. However, information users in the University of the Philippines perceive that the librarians/information providers still need further training in order to fully understand the potential of such technologies and increase their competence in manipulating these sophisticated tools. Training should augment their experience and should specifically develop their expertise in information retrieval.

The problems we encounter, as with other Third World libraries, can likewise be alleviated if a budget can be allocated to sustain continuous on-the-job training for the librarians/information providers, because the pace of technological change is at such rate that computers have an average life span of three to five years and the electronic environment changes rapidly.

 

54

 

References

Hilotin, C.I. 1994. Computer networking need not be expensive. TODAY, 22 Apr. 1994:13.

Keylard, M. 1993. CD-ROM implementation in developing countries: impacts and pitfalls. IFLA Journal 19(1):35-49.

Kinnucan, M.T. 1993. Demand for document delivery and interlibrary loan in academic settings. Library and Information Science Research 15(4):355-374.

Magno, L 1994. Telecoms link to Internet that will flood RP with trade info being finalized. TODAY, 8 Apr. 1994:13

Stueart, RD. 1982. Libraries: a new role. In: Books, libraries and electronics: essays on the future of written communication (ed. by E. Sigel, et al.). New York: Knowledge Industry Publications, Inc.

Valentin, R. 1995. Information: a global resource. IDRC Reports 23(4):4-6.

 

55

 

JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617

HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/

Plagiarism Detection Using Graph-Based

Representation

1Ahmed Hamza Osman, 2Naomie Salim, 3Mohammed Salem Binwahlan

1Faculty of Computer Science, International University of Africa, Sudan

2Faculty of Computer Science and Information Systems, Universiti Teknologi Malaysia

3Faculty of Applied Sciences, Hadhramout University of Science & Technology, Yemen

1, 2, 3 81310, Skudai, Johor, Malaysia

Tel +60147747409/+607 5532208 Fax +607 5532210

Abstract—Plagiarism of material from the Internet is a widespread and growing problem. Several methods used to detect the plagiarism and similarity between the source document and suspected documents such as fingerprint based on character or n-gram. In this paper, we discussed a new method to detect the plagiarism based on graph re-presentation; however, Preprocessing for each document is required such as breaking down the document into its con¬stituent sentences. Segmentation of each sentence into separated terms and stop word removal. We build the graph by grouping each sentence terms in one node, the resulted nodes are connected to each other based on order of sentence within the document, all nodes in graph are also connected to top level node” Topic Signature “. Topic signature node is formed by extracting the concepts of each sentence terms and grouping them in such node. The main advantage of the proposed method is the topic signature which is main entry for the graph is used as quick guide to the relevant nodes. which should be considered for the comparison between source documents and suspected one. We believe the proposed method can achieve a good performance in terms of effectiveness and efficiency. 

— Plagiarism detection, graph representation, concept extraction, topic signature

- - - - - - - - - - 

 

P

LAGIRAISM is one of the forms of misuse of academ¬ic activities has increased rapidly in the quick and easy access to data and information through electron¬ic documents and the Internet, and when we talk about plagiarism, we mean the text written by others where they are re-adjust the text to format by adding or deleting without any citation or reference.

There are many types of plagiarism, such as copy and paste, which is the most common, redrafting or paraph-rasing of the text, plagiarism of the idea, plagiarism through translation from one language to another and many other methods that use plagiarism. Plagiarism is a serious problem in computer science. This is partly due to the ease with which electronic assignments may be co-pied, and to the difficulty in detecting similar as-signments in a sufficiently large class. In addition, students are becoming more comfortable with cheating. A recent study found that 70% of students admit to some plagiarism, with about half being guilty of a se 

• Ahmed Hamza Osman. is with the Faculty of Computer Science, Interna-tional University of Africa, Sudan.

• Naomie Salim. Faculty of Computer Science and Information Systems, Universiti Teknologi Malaysia,malsysia.

• Mohammed Salem Binwahlan. is with the Faculty of Applied Sciences, Hadhramout University of Science & Technology, Yemen.

 

rious cheating offense on a written assignment. In addition, 40% of students admit to using the “cut-and-paste” approach when completing their assign¬ments [1]. The key and main issue in plagiarism detec¬tion field is how to differentiate between plagiarized doc¬ument and non-plagiarized document in effective and efficient way.

The current methods of plagiarism detection relay on the comparison of small text unit such as character, n-gram, chunk or terms. Suppose we have a document contents ten sentence, each sentence contains five terms and each term consists of at least one character. The consideration of small text unit (character) for detecting of similarity between the original document and suspected document lead to a huge number of comparisons. In this paper, we propose a new method for plagiarism detection. The pro-posed method is graph-based, where each document is represented as graph. One node represents one sentence. Top level node is different node where it contains the concepts of terms in the document. Such node is called topic signature. The main advantage of proposed method is the topic signature which is the main entry for the graph is used as quick guide to the relevant nodes, which should be considered for the comparison between source documents and suspected one.

 

JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617

HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ 37

 

This paper is structured as follows. Section 2 introduces the current state of the studies on plagiarism detection. While Section 3 presents an overview of the graph based representation, Section 4 describes the concepts extraction for the sentence. Contribution and comparison methods are reported in Section 5. Finally, Section 6 concludes this paper.

In plagiarism detection, a correct selection of text features in order to discriminate plagiarised from non-plagiarised documents is a key aspect. [2] Has delimited a set of fea¬tures which can be used in order to find plagiarism cases such as changes in the vocabulary, amount of similarity among texts or frequency of words. This type of features has produced diferent approaches to this task. Substan¬tive plagiarism analysis [3] is a different task from pla¬giarism detection with reference. It captures the style across a suspected document in order to find fragments that are plagiarism candidates. This approach saves the cost of the comparison process, but it does not give any hint about the possible source of the potentially plagia¬rised text fragments. In those cases where a reference cor¬pus is considered, the search process has been based on different features. [4] Considers text comparison based on word n-grams. The reference, as well as the suspected text, is split into trigrams, composing two sets which are compared. The amount of common trigrams is considered in order to detect potential plagiarism cases. [5] Considers the sentence as the comparison unit in order to compare local similarity. It differentiates among exact copy of sen-tences, word insertion, word removal and rewording on the basis of a Wordnet-based word expansion process.

some authors [6] define plagiarism as “unacknow¬ledged copying of documents or programs” that can “oc¬cur in many contexts: in industry a company may seek competitive advantage; in academia academics may seek to publish their research in advance of their colleagues.” Most empirical study and analysis has been undertaken by the academic community to deal with student plagiarism, although methods of detection have found their way into the commercial world, e.g. Measuring software reuse and identifying reused code (see, e.g. [7]).

There are several schemes to characterize documents be-fore applying one of the plagiarism detection techniques. Some document descriptors such as Character-based representation, the simplest form, in which documents are represented as a sequence of characters with ignoring spaces between words, periods (full stops) between statements and lines. Also Word-based representation, in which documents are represented as a collection of words with ignoring periods (full stops) between statements and lines. Moreover Phrase-based representation, in which a phrase (part of a statement) is used as a unit of compari¬son. For example, 3-word phrase or so-called trigrams can be used as a comparison unit. and Sentence-based 

 

representation, in which documents are segmented into statements using periods (full stop) as a statement-end indicator. Although Paragraphed-based representa¬tion, in which documents are described as a collection of paragraphs or passages.

There are several techniques have been developed or adapted for plagiarism detection in natural language documents. They can be classified into four main ap-proaches. The first technique is Fingerprint Matching [8][9][10] which involves the process of scanning and examining the fingerprints of two documents in order to detect plagiarism. Then, Clustering [11][10] that uses specific words (or keywords) to find similar clusters between documents Fingerprinting techniques mostly rely on the use of K-grams [12] because the process of fingerprinting divides the document into grams of certain length k. Then, the fingerprints of two documents can be compared in order to detect plagiarism. It can, therefore, be classified fingerprints into three categories: character-based fingerprints, phrase-based fingerprints and statement-based fingerprints. The early fingerprint-ing technique uses sequence of characters to form the fingerprint for the whole document.

Some authors refer about the tools in plagiarism detection [13] which are currently particularly popular and describe their main features in what follows.

www.plagiarism.org

Turnitin: This is a product from iParadigms. It is a web based service. Detection and processing is done remotely. The user uploads the suspected document to the system database. The system creates a complete fingerprint of the document and stores it. Proprietary algorithms are used to query the three main sources: one is the cur¬rent and extensively indexed archive of Internet with approximately 4.5 billion pages, books and journals in the ProQuestTM database; and 10 million documents already submitted to the Turnitin database.

www.urkund.com

Urkund: Another server based plagiarism detection web service which offers an integrated and automated solution for plagiarism detection. It utilizes standard email systems for submission of documents and viewing results. This tool also claims to search through all available online sources giving priority to educational and scandinavian origin. This system claims to process 300 different types of document submissions.

www.copycatchgold.com

Copycatch: A client based tool used to compare lo¬cally available databases of documents. It offers ‘gold’ and ‘campus versions’ , giving comparison capabilities for large number of local r sources. It also offers a web version which extends the capabilities of plagiar 

 

JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617

HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ 38

 

ism detection across the internet using the Goggle API. www.plagiarism.phys.virginia.edu

WCopyfind: An open source tool for detecting words or phrases of defined length within a local repository of documents . The product is being modified to extend searching capabilities across the internet net using the Google API at ACT labs10.

www.canexus.com

Eve2 (Essay Verification Engine): This tool works at the client side and uses it own internet search mechanism to find out about plagiarized contents in a suspected document.

http://www.plagiarism.com

GPSP - Glatt Plagiarism Screening Program: This soft-ware works locally and uses an approach to plagiarism detection that differs from previously mentioned ser-vices. GPSP detection is based on writing styles and pa terns. The author of a suspected submission has to go through a test of filling blank spaces in the writing. The number of correctly filled spaces and the time taken for completion of the test provides the hypothesis of pla-giarism guilt or innocence.

www.cs.berkeley.edu

MOSS - a Measure of Software Similarity: MOSS In¬ternet service “accepts batches of documents and re¬turns a set of HTML pages showing where significant sections of a pair of documents are very similar ” [14]. The service specializes in detecting plagiarism in C, C++, Java, Pascal, Ada, ML, Lisp, or Scheme programs.

www.ipd.uni‐karlsruhe.de

JPlag: Another internet based service which is used to detect similarities among program source codes. Users upload the files to be compared and the system presents a report identifying matches. JPlag does programming lan¬guage syntax and structure aware analysis to find results.

Graph-based method is introduced, designed especial¬ly for web document representation [15]. “ The main advantage of graph-based techniques is that they allow keeping the inherent structural information of the original document. Before describing the graph-based methodol¬ogy, the definition of a graph, subgraph and graph iso¬morphism should be given. A graph G is a 4-tuple: G= (V,E,α,β), where V is a set of nodes (vertices), E ك V×V is a

set of edges connecting the nodes, á : V  v is a

function labeling the nodes, and â : V×V  e is a

function labeling the edges (v and e being the sets of labels that can appear on the nodes and edges, re¬ 

 

spectively). For brevity, we may refer to G as G= (V,

E) by omitting the labeling functions. A graph

G1=(V1,E1,α1,β1) is a subgraph of a graph G2=(V2,E2,α2,β2), denoted G1 ك G2, if V1 ك V2, E1 ك E2 0 (V1 × V1), á1(x) = á2(x) ׊ xאV1 and â1(x, y) = â2(x, y) ׊ (x, y) א E1. Conversely, graph G2 is also called a supergraph of G1. All graph representations proposed in [ 15] are based on the adjacency of terms in an HTML document. Under the standard method[16] each unique term (word) appearing in the document, except for stop words such as “the”, “of”, and “and” which convey little information, becomes a vertex in the graph representing that document. Each node is labeled with the term it represents. Note that we create only a single vertex for each word even if a word appears more than once in the text to build the terms graph in the sentence, and we create also a single vertex for each sentence this vertex involved graph of terms. Also Under the n-distance representation, there is a user-provided parame-ter.Instead of considering only terms immediately follow-ing a given term in a web document. For example[16], if we had the following text on a web page, “AAABBBCCCDDD”, then we would have an edge from term AAA to term BBB labeled with a 1, an edge from term AAA to term CCC labeled 2, and so on. Similar to n-distance, we also have the fourth graph representation, n-simple distance. This is identical to n-distance, but the edges are not labeled, which means we only know that the distance between two connected terms is not more than n. frequency representation model is a type of graph representation too. Each node and edge are labeled with an additional frequency measure. For nodes this indicates how many times the associated term appeared in the web document; for edges, this indicates the number of times the two connected terms appeared adjacent to each other in the specified order”. We discussed the re-presentation of graphs in this paper in section 5.

Concept identification is a common to applications such as ontology learning, glossary extraction and keyword extraction. These applications have different definitions for concept, hence different methods. Pre¬vious methods start from the idea that concepts can be found as word or phrases contained in sentences, which are then divided into smaller phrases in one of two ways: Using grammatical or syntactical information. The former can be found in ontology learning [17], glossary extraction [18] and information retrieval systems [19]. Using a shallow grammar parser, an entire sentence is parsed into a grammatical tree, which classifies sub-phrases as noun or verb phrases, noun phrases are selected as concepts. The syntactical information division of sentences uses punctuation or conjunctions to separate phrases within a sentence, all these phrases are concepts. This approach can be found in keyword extraction systems [20].

The idea that was used in this paper refers to extrac-tion the general concepts from the sentences by any me 

 

JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617

HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ 39

 

 

Fig. 1. Represent sentences in the nodes and extract the concepts from the sentences and the weight between the nodes refer to similari-ty between the sentences, T1, T2,... ,Tn represent the terms.

thod used to extraction. We suggest the same ideas used in the extraction of concepts to be applied in the method that we have proposed.

In this section we discuss the new method for plagiarism detection based on the graph representation. The method relies on a number of steps, first we break down the doc-ument into its constituent sentences. Preprocessing for each document is required such as segmentation of each sentence into separated terms and stop word removal and. the stemming process is applied on the sentence, then we represent the sentences in the form of a nodes related to edge on the order of the sentence within the document. Where each node contains a one sentence of the document. To represent the terms of each sentence as graph we use the method mentioned in section 3. Where each node in the graph contains one term. The node are connected to each other according to order of term posi-tion in the sentence. The whole document consists of a number of nodes determined by the number of sentence in the document. Each node is a graph which represents a sentence. The concepts of node terms are extracted and used for calculating the similarity between each pair of nodes using e.q. (1).

Where Si is sentence 1 and Si+1 is sentence following the sentence1 ,Csi is a number of concepts in sentence I and Csi+1 is a number of concepts in sentence i+1, W is a weight or similarity between si and si+1.

After getting all the concepts of sentences those concepts 

 

grouped in one node called the topic signature, this node inked to each node in the graph and then the similarity between the topic signature and each node is computes separately, We calculate the similarity between the node of the topic signature and the other nodes based on shared concepts using the following equation:

W

Topic. Signature

s

Where Csi is a number of concepts in the sentence I, C is the number of concepts in the topic signature node.

Topic signature node is formed by extracting the concepts of each sentence terms and grouping them in such node. The main advantage of the proposed method is the topic signature which is main entry for the graph is used as quick guide to the relevant nodes. which should be con¬sidered for the comparison between source documents and suspected one. For example, if there is a matching between concept1 on the suspected document and con¬cept 1 in the original document, we go directly to the nodes that containing of concept1 in both original docu¬ment and the suspected document and we ignore all the remaining sentences. In the case of full matching of topic signature of original document with topic signature of suspected document, we will face the problem of the huge number of comparisons which is taken as disadvan¬tage in current methods. To avoid such problem, only we compare the most important nodes. To determine those important nodes we link each node with all the nodes, However, each node has a number of in-links and out-links then compute the similarity of each sentence with the rest of the sentences in the document. Based on the similarities, we extract the highest degree of similarity between the nodes which define the most important nodes and ignore the rest. To calculate the degree of simi¬larity of the node with other nodes using the following equations

c c

W  (  )  si n sk out link si sk

csk

c nc

W ( si sk 

i si sk )

si

Where is a sentence I, Csi is a number of concepts in the sentence I, Csk is a number of concepts in each sentence in the graph.

 

JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617

HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ 40

In this paper we have considered the problem of pla-giarism detection, one of the most publicized forms of text reuse around us today. In particular, we have focused on plagiarism detection using graph based document representation. We have discussed various approaches of plagiarism detection. To date there are few resources which specifically address the task of plagiarism detection.

The proposed method to detect the plagiarism based on graph representation required break down the document into its constituent sentences and the graph building by grouping each sentence terms in one node, the resulted nodes are connected to each other based on order of sen¬tence within the document, all nodes in graph are also connected to top level node” Topic Signature “.  

Fig. 2. Topic Signature formulation for concepts Sentence as Graph

 

Topic signature is a main entry for the graph is used as quick guide to the relevant nodes. Which should be considered for the comparison between source documents and suspected one.

The proposed method based on graph representations contributed by increase the efficiency and reduce the huge a number of matching process.

[1] D. McCabe. Levels of Cheating and Plagiarism Remain High. Center for Academic Integrity, Duke University, 2005. Website: http://academicintegrity.org 

[2] Clough. P.: Plagiarism in Natural and Programming Lan-guages: an Overview of Current Tools and Technologies. Research Memoranda: CS-00-05, Department of Computer Science. University of Sheffield, UK (2000)

[3] Meyer zu Eissen, S., Stein, B.: Intrinsic plagiarism detec-tion. In: Lalmas, M., Mac- Farlane, A., R¨uger, S.M., Tom-bros, A., Tsikrika, T., Yavlinsky, A. (eds.) ECIR 2006. LNCS, vol. 3936, pp. 565–569. Springer, Heidelberg (2006)

[4] Lyon, C., Barrett, R., Malcolm, J.: A Theoretical Basis to the

Automated Detection of Copying Between Texts, and its Practical Implementation in the Ferret Plagiarism and Col¬lusion Detector. In: Plagiarism: Prevention, Practice and Policies Conference, Newcastle, UK (2004)

[5] Kang, N., Gelbukh, A., Han, S.-Y.: PPChecker: Plagiarism pattern checker in document copy detection. In: Sojka, P., Kopeˇcek, I., Pala, K. (eds.) TSD 2006. LNCS (LNAI), vol. 4188, pp. 661–667. Springer, Heidelberg (2006)

[6] Joy, M. and Luck, M. (1999), Plagiarism in Programming Assignments, IEEE Transactions of Education, Vol. 42(2), 129-133.

[7] Hislop, G. W. (1998), Analyzing existing software for soft-ware reuse, Journal of Systems and Software, Vol. 41, 33-40.

[8] Heintze, N. (1996). Scalable document fingerprinting. Pa-per presented at the Second USENIX Workshop on Elec-tronic Commerce.

 

[9] Lyon, C., Malcolm, J. A., & Dickerson, R. G. (2001). Detect¬ing short passages of similar text in large document collec¬tions. Paper presented at the Conference on Empirical Me¬thods in Natural Language Processing.

[10] Yerra, R., & Ng, Y.-K. (2005). ASentence-Based Copy Detec¬tion Approach for Web Documents. In Fuzzy Systems and Knowledge Discovery (pp. 557-570).

[11] Antonio, S., Hong Va, L., & Rynson, W. H. L. (1997). CHECK: a document plagiarism detection system. Paper presented at the Proceedings of the 1997 ACM symposium on applied computing.

[12] Manuel, Z., Marco, F., Massimo, M., & Alessandro, P.

(2006). Plagiarism Detection through Multilevel Text

Comparison. Paper presented at the Second Interna¬tional Conference on Automated Production of Cross Me¬dia Content for Multi-Channel Distribution.

[13] Maurer, H., F. Kappe, B. Zaka. Plagiarism – A Survey. Journal of Universal Computer Sciences, vol. 12, no. 8, pp. 1050 – 1084, 2006.

[14] S. Schleimer, D. S. Wilkerson, and A. Aiken. Winnowing: local algorithms for document fingerprinting. In SIGMOD: Proceedings of the 2003

[15] A. Schenker, H. Bunke, M. Last, and A. Kandel, "Graph-Theoretic Techniques for Web Content Mining", Series in Machine Perception and Artificial Intelligence, 62, World Scientific, 2005.

[16] A.SCHENKER, M. LAST, H. BUNKE3, AND

A.KANDEL,“CLASSIFICATION OF WEB DOCUMENTS USING GRAPH MATCHING”

[17] R. Navigli and P. Velardi, “Learning Domain Ontologies from Document Warehouses and Dedicated Web Sites," Computational Linguistics, vol. 30, pp. 151--179, 2004.

[18] D. Bourigault and C. Jacquemin, "Term extraction + term clustering: An integrated platform for computer-aided terminology," in EACL, 1999.

[19] I. Bichindaritz and S. Akkineni, “Concept Mining for In-dexing Medical Literature," Lecture Notes in Computer Science, vol. 3587, pp. 682--692, 2005.

[20] I. H. Witten, G. W. Paynter, E. Frank, C. Gutwin, and C. G.

 

JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617

HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ 41

Nevill-Manning, "KEA: practical automatic key phrase extraction,” in Fourth ACM conference on Digital libraries, 1999.

is a PHD student in Universiti Teknologi Malay-sia. Presently working as a lecturer in the Faculty of Computer Science in International University of Africa(IUA) in Sudan .I have bachelor degree in Computer Science from IUA in 2004. I received master degree in Computer Science from Sudan University of science and technology in 2008. I current research interest includes Information Retrieval, Database, and Data mining.

is an Associate Professor presently working as a Deputy Dean of Research & Postgraduate Studies in the Faculty of Computer Science and Information System in Universiti Teknologi Malaysia. She received her bachelor degree in Computer Science from Universiti Teknologi Malaysia in 1989. She received her master degree in Computer Science from University of West Michigan in 1992. In 2002, she received her Ph.D (Computational Informatics) from University of Sheffield, United Kingdom. Her current research interest includes Information Retrieval, Distributed Database and Chemoinformatic.

received his B.Sc. dgree in

Computer Science from Hadhramout University of Science and

Technology, Yemen in 2000. He received his Master degree

from Universiti Teknologi Malaysia in 2006. He is currently

with Hadhramout University of Science and Technology as lecturer and pursuing Ph.D degree in the Faculty of Computer Science and Information System, Universiti Teknologi Malay¬sia. His current research interest includes Information Retriev¬al, Text Summarization and Soft Co puting.

 

Roman Business Law

Jean-Jacques Aubert, Université de Neuchâtel

Historians of the Roman economy seem to agree that during the period 200 BC – AD 200 the Gross Domestic Product of the overall empire grew, however moderately. They explain this phenomenon partly as the result of a reduction of transportation costs and of ‘transaction costs,’ namely the sum of the costs of looking for opportunities for exchanging goods and services, of reaching agreements between parties through contracts, and of

enforcing such transactions.1 In other words, this reduction was due to, among other causes, the development of a common legal system, ‘especially in the field of commercial

law.’2 This statement begs the question: What is ‘Roman commercial law’? Where does it come from? How did it develop? And how does it fit in the wider field of Roman law? This chapter will attempt to provide some answers.

Law of commerce, commercial law, business law

A standard handbook of Roman law unambiguously states at the outset: ‘Ein

besonderes Handelsrecht haben die Römer daneben nicht ausgebildet.’3 However, there is no dearth of books and articles entitled, or explicitly dealing with, Diritto commerciale romano

or L’histoire du droit commercial romain.4 Behind this apparent contradiction lurk both a question of definition and a recurrent and on-going scholarly debate concerning the way to

approach Roman legal institutions governing trade and other economic activities.5 ‘Commercial law’ (Handelsrecht), with its alleged specificity (‘besonderes’) and independent status (‘daneben’), and its modern scion, ‘business law’ (droit des affaires), will serve as heuristic instruments to evaluate the usefulness, sophistication, and shortcomings of the Roman law of commerce.6

According to a strict definition, ‘commercial law’ is a set of legal rules originating with merchants, designed for merchants, and enforced – partly at least – by merchants. Scholars looking for Roman ‘commercial law’ focus on sources of law, legal interpretation, practical

1 As suggested by the New Institutional Economics (NIE) school of thought, cf. Thomas, cf. North – Thomas (1973); North (1981) and (1991); and Malanima (2009), 7-8.

2 Lo Cascio (2007), 619 and 626 (quotation); Harris (2003), 282 and 285; Silver (2007), 192 (quoting A. Wacke and H.W. Pleket), 210-11; and 217-20; and Terpstra (2008).

3 Kaser (1971) 474 (‘The Romans did not develop a specific, independent commercial law.’)

4 Cf. most recently Cerami – Petrucci (2010). Cf. also Di Porto (1997), citing Carnazza (1891) (non vidi); and Bianchini (1989) and (2007), with some caution. Labruna (1994) coins the term ‘diritto mercantile’. Others speak only of ‘istituti commerciali del/nel diritto romano’: cf. Fadda (1903/1987); Cogliolo (1922); and Földi (2001), 85.

5 The debate started anew in Italy in the late 1980s, cf. Bianchini (1989) and (2007), and Di Porto (1997). It was subsequently carried on at various conferences, e.g. SIDA in Rotterdam (2001), cf. Gerkens (2003); Circolo Toscano ‘Ugo Coli’ at the Certosa di Pontignano (12-14 Jan. 2006); and the Fourth OxREP conference, 1-4 October 2009, cf. Sirks (forthcoming).

6 For surveys in English, cf. Crook (1967), ch. 7, 206-49; and Johnston (1999), ch. 5, 77-111.

 

application, and jurisdiction. The former two are quite accessible through extant sources, the latter two mostly blurred for a lack of them.

Historically, the definition proposed above applies to the law developed in Western Europe, mostly Italy and France, during the Middle Ages. Thus, ‘commercial law’ (ius mercatorum, lex mercatura, Law Merchant) was meant to be more pragmatic and flexible, less bookish, and less dominated by scholars than Roman and Canon laws. Its purpose was to satisfy the needs of commerce, facilitating transactions, expediting proceedings through separate jurisdictions and procedures, and transcending the limits of national legal systems as an early form of international law. Its provisions unavoidably reflect the concerns and interests of an identifiable socio-economic class of people, i.e. traders. ‘Commercial law’ was a part of private law, and its status with regard to civil law has been fluctuating throughout history: distinct at first, it tended to merge and be absorbed by the latter, as is the case in Swiss or Italian law, eventually becoming a subfield of the law of obligations. Alternatively, ‘commercial law’ opened up to other fields of law and became more inclusive in terms of both people ruled by it and types of issues and transactions dealt with it, thus evolving into ‘business law’ (cf. below).7

In the Middle Ages, cases of ‘commercial law’ were heard in special courts. As litigation frequently involved people of various national and social origins, judges based their decision on a mix of mercantile codes and usages, while paying particular attention to good faith. The formalism of legal procedure was somewhat relaxed, and judicial decisions

were rendered on the basis of the consideration of facts rather than legal technicalities.8

Originally, ‘commercial law’ was mostly maritime law, with an emphasis on contracts related to sales, transportation, and money-lending. With time, the nature of trade became more diverse and more complex. ‘Commercial law’ also applied to land-based trade (in the context of fairs and markets), including production, storage, and distribution, and to people involved in any economic activity. A final development even took consumption and consumer protection into consideration. A standard modern treatise of (British) commercial law deals with property, contract (mostly sales and partnerships), agency, payment instruments and systems, financing, insolvency (in connection with company law),

and the resolution of commercial disputes through litigation or arbitration.9

While ‘commercial law’ was exclusively concerned with identifiable (i.e. registered) businesses, and exclusively applied to business transactions performed by well-defined groups of people, traders and professional businessmen, such restrictions came to be seen as counterproductive, and called for adjustment. According to recent trends ‘commercial law’ evolved into ‘business law,’ disregarding the specific status of both structures and people, and combining elements of both private and public law. Legal practitioners thus

7 Merryman (1985), 12-13, 98-100, and 146-47.

8 Goode (2004), 3-6.

9 Goode (2004).

2

 

acquired a transversal competence allowing them to deal with all kinds of issues relating to the multifarious aspects of commercial life.10

Because ‘business law’ is more inclusive than ‘commercial law,’ and because of the reciprocal ‘civilizing’ of commercial law and ‘commercializing’ of civil law which resulted in the development of ‘business law,’ it seems relevant to look at Roman legal institutions, both private and public, in their historical development from the earliest time until the period of classical law in order to evaluate their commercial relevance, usefulness, and adequacy, in comparison with the trappings of both later ‘commercial law’ and ‘business law.’ The following survey will show that while the appellation of Roman ‘commercial law’ is unsustainable, the Roman law of commerce shares many features with modern ‘business law,’ features which extend far beyond the scope and limits of the Roman law of obligations.

Toward a history of Roman business law

‘Buying and selling originate with exchange or barter.’ These are the opening words of the title on sale in Justinian’s Digest, excerpted from Paul’s commentary on the praetorian edict (33 ad ed., D. 18.1.1 pr.). Trade certainly existed before extant sources reveal how Roman law dealt with it. In a pre-monetized society, the exchange of goods and services was assumedly based on barter or exchange (gr. amoibè/lat. permutatio).11 In spite of the

symmetrical feature of barter,12 both parties being equal with regard to the uncertainty about the quality of goods to be exchanged, barter was viewed by classical jurists as an impediment to trade, because of the difficulty to have demand meet supply. To be sure, some commodities (such as cattle, metals, slaves, staples, etc.) were deemed universally desirable and were used as monetary instruments at a very early date. The consensus on a constant medium of exchange eventually took the form of coinage. While relying on bronze bullion (aes rude/grave), the Romans started using Greek coinage by the fifth or fourth century BC, and coined their own by the early third. Barter undoubtedly gave rise to disputes since Rome’s earlier period, but there is no trace of any litigation connected with it. It is possible that barter was not legally recognized before a much later period and that social control was sufficient to settle such disputes. Besides, non-monetary commercial exchanges must have existed throughout Roman history, especially wherever and whenever currency was a scarce commodity, and monetization an unfamiliar abstraction. Even though the geographer Strabo, active in the Augustan period, associates barter with backwardness and uncivilized ways of life typical of marginal, unassimilated tribes,13 it must have been a Roman reality all along.

10 Champaud (1994); Legeais (2003); Lucas (2005); Kelly – Holmes – Hayward (2005). For the combination of private and public laws, cf. Goode (2004), 10; and Lucas (2005), 8.

11 Meylan (1959). Kudlien (2001) rightly points out that the word permutatio has various meanings.

12 Morley (2007) 59, as opposed to sale, where the seller had the advantage of knowing the quality of the goods on sale.

13 Strabo 3.3.7 (about Lusitanian mountain-dwellers, trading through exchange/amoibè or bullion); and 7.5.5 (about Dalmatians). Cf. de Churruca (2001).

 

3

 

Classical Roman jurists knew of the practice of barter and dealt with it rather marginally. In the first century AD Sabinus and Cassius thought of it as equivalent to sale, while Nerva and Proculus disagreed with them (Paul [33 ad ed.] D. 18.1.1.1). The mid-second-century jurist Gaius (Inst. 3.141), following Sabinus, underlines its antiquity by citing Homer (Il. 7.472-475) and reports earlier disputes concerning its contractual status. Around the time of Trajan, Sextus Pedius and Aristo, both cited and followed by Paul a century later ([33 ad ed.] D. 19.4.1.3 and [5 ad Plautium] 19.4.2), address marginal issues by analogy with the consensual contract of sale (emptio-venditio). Paul – the only classical jurist whose works are excerpted in the title of the Digest dealing with barter (D. 19.4, De rerum permutatione) – wonders about the nature of the obligation (re) arising from such transaction, buyer (emptor) and seller (venditor), price (pretium) and good (merx) being undistinguishable from one another, with negative consequences in case of non-delivery or eviction.14 It is only in the late classical period that barter was promoted from the status of ‘unenforceable pactum’ to that of so-called ‘innominate real contract.’ The Justinianic Code preserves several imperial constitutions dated to the mid- and late third century AD on the subject (CI. 4.64.1-8), suggesting that as payments in kind may have increased when the Roman monetary system was in shamble, the need to regulate this type of commercial exchanges may have become more urgent.

The history of barter in Roman law serves as a reminder that all economic transactions were not necessarily sanctioned by law.15 Looking at commerce in general, the first (i.e. both earliest and most prominent) problem to be dealt with is legal recognition and, consequently, jurisdiction.

Ius comm ercii

Among the first policies enacted by the Republican state after the revolution of 509 BC, an important step consisted in establishing and defining (commercial) contacts with neighbors, both immediate (Latins and other Italic people in the region) and farther ones (Etruscans, Greeks, and Carthaginians, to name only the most important ones). Polybius (3.22-27) records the content of a series of treaties between Romans and Carthaginians. In the first treaty (ca. 509-507), it is agreed upon that trade carried out by the Romans in Sardinia, Sicily, and Africa – all Carthaginian territories at the time – should be strictly controlled and guaranteed by the state. Transactions must be concluded through an auctioneer (kèrux) and a scribe (grammateus), both of whom engage the Carthaginian state’s good faith (pistis) toward sellers (3.22). Reciprocity is not mentioned. This provision sounds like a protection against piracy, ransoming, or extortion, as the distinction between such practices and trade is sometimes blurred. The second treaty (ca. 348, cf. Liv. 7.27) explicitly forbids piracy and opens up trade in Sicily and Africa for the Romans and at Rome for the Carthaginians, where traders of each nation had the same rights as the natives (3.24). This means that the exchange of goods between Romans and Carthaginians had the same legal

14 D. 19.4.1-2 (with Mommsen’s correction of the provenance). Cf. also id. (ibid.) D. 18.1.1 pr.-1; and Ulp. (1 ad ed. aed. cur.) D. 21.1.19.5, possibly itp. Cf. Zimmermann (1990), 250-52; 532-37; Johnston (1999), 78-79.

15 Aubert (2007b), for such a situation in fourth-century Athens.

 

4

 

validity in either place as between fellow countrymen. A third (or fourth?, Liv. 9.43) treaty, dated ca. 278, reasserts these dispositions. Subsequent treaties, while redefining areas of respective power and influence, do not question the basic trade agreement, the text of which, interestingly, was engraved on bronze tablets and preserved in the aediles’ office (3.26).16

Whereas Polybius does not enter into detail about the legal aspects of such arrangements for international trade, Dionysius of Halicarnassus reports (6.95), in the context of the so-called foedus Cassianum, a treaty concluded ca. 493 with neighboring Latin communities, that contracts between Romans and Latins would be enforced in courts, within ten days, wherever they had been concluded. This provision implies that in any court of law Romans and Latins would enjoy an identical legal standing, with equal protection from the law recognized and enforced by the respective courts. This would have excluded the potential contradictions attached to competing legal systems and inaugurated a form of international law (ius gentium).17 Accessorily, plaintiffs are guaranteed a speedy trial, facilitated by the fact that discrepancies between legal systems should be irrelevant. This arrangement will be known subsequently to the Romans as commercium/ius commercii (Tit. Ulp. 19.5), namely the right to make formal contracts, to acquire property, and to resort to courts, according to Roman law and procedure.18 It was, or became, part of a larger package (isopoliteia) including the right to intermarry (conubium) and to participate in civic life (suffragium). This interpretation is based on the terms of the settlement of 338 BC, whereby Latins would be deprived of various rights they previously enjoyed (Liv. 8.14.10). It is quite possible that commercium allowing Latins to benefit from the protection of Roman law was more than the rights secured through the Romano-Carthaginian treaties.19 It is also likely that the scope of commercium was limited in comparison with Roman citizenship and did not come close to extending to those privileged foreigners (peregrini) a legal protection equal to that enjoyed by Roman citizens.20

The issue of legal recognition of commercial transactions concluded with foreigners was taken up in the first Roman ‘codification’ of law known as the Law of the Twelve Tables, dated ca. 451-450 BC. Foreigners (hostes, cf. Cic., Off. 1.37), assumedly with commercium, can have their day in court (2.2, Crawford), possibly with some degree of priority over other cases with respect to international treaties,21 but transfer of ownership cannot occur through long-term prescription (usucapio), as the retention of eminent title (auctoritas) with regards to, or against the claim of, foreigners is not limited to one or two

16 Ferenczy (1969); Scardigli (1991); Nörr (2005), esp. 171-77, where the author sees in Pol. 3.22.8 a reference to two common forms of sales (auctioneering and written contract), and suggests a possible interpretation for the term telos (auctoritas) as the effect of the contract.

17 Nörr (2005) 183-84, with reference to Wolff (1979).

18 Mayer-Maly (2003); Nörr (2005); and Minaud (2011), ch. 1 (§§ 19-70, esp. 35-36).

19 Capogrossi Colognesi (1994); Kremer (2005).

20 Kaser (1953), suggesting that mancipatio, not in iure cessio, was made available to foreigners as a formal way to convey property.

21 Kremer (2005), 197-203.

 

5

 

years (6.3), but should be everlasting (6.4).22 This means that foreigners can only acquire ownership through formal conveyance, such as mancipatio (Tit. Ulp. 19.4), and that the seller/transferor would have to protect such a buyer/transferee against eviction by a third party for an unlimited period of time.

In this regard, the Law of the Twelve Tables contains some dispositions offering potential for innovation. Take, for example, the law of contracts (6.1): ‘When someone shall perform a nexum or a mancipatio, the ius will be defined by what the tongue has pronounced.’ Nexum is an early form of loan (Varro, LL 7.105), performed, like mancipatio, ‘by means of bronze and scale’

(per aes et libram) and guaranteed by the pledge of the very person of the debtor. Regarded as unduly risky and anti-social in its consequences, it was abolished in the late fourth century by a lex Poetelia Papiria (326 or 313 BC). Mancipatio proved a more durable institution. The provision of the Twelve Tables introduces a verbal dimension to the formal act, allowing the parties to specify the terms of the contract to be concluded. Originally the solemn utterance before witnesses called nuncupatio may have been more or less fixed, prescribed words being imposed on the parties. The sheer fact that the same ritual per aes et libram was performed in widely different contexts, such as the making of a will or a donation, the conveyance of property, the constitution of a dowry or servitude, the emancipation of a dependent, or the contracting of a loan, points to a large spectrum of required statements. In pre-classical Roman law, contractual obligations became overwhelmingly verbal, the so-called stipulatio (and its variants, sponsio, promissio, cautio, etc.) being so flexible – and liable to become increasingly so – as to adequately address most social and economic needs. Nuncupatio and stipulatio share the faculty of clarifying intentions in any legal situation. In both cases only one party’s intention is clarified.

It is remarkable that the Twelve Tables, for all of their preserved or reconstructed provisions (between 88 and 109), have little to say about commerce.23 Provisions 3.5 and 3.6 incidentally allude to periodic market-days (nundinae). Some other clauses are pregnant with important features of later legal developments in the law of commerce, such as the civil liability of masters for the (wrong)doings of their dependents (8.2 and 12.2, noxa), deceit (8.10, fraus), and malice aforethought (8.9, dolus malus). Much of the law, however, is concerned with criminal law, police regulation, and civil procedure. This latter field, with its reliance on the role of magistrates and despite its formalism, would prove instrumental in the development of much of the Roman law of commerce over the next three centuries or so (mid-fifth to late-second c. BC).

In the archaic and mid-republican periods, civil litigation falls within the scope of one of five actions of the law (legis actiones), or general remedies granted by a magistrate endowed with jurisdiction based on imperium. These remedies ranged from the taking of a pledge (pignoris capio) or personal execution (manus iniectio) to the initiating of a judicial procedure in front of the magistrate (iudicis arbitrive postulatio), on the basis of a specific

22 Pace Kremer (2005), 203-06; and Humbert (2005), 393-97.

23 Aubert (2004), 164-65. As W.V. Harris suggested to me, those archaic rules may have been completely superseded by the time the law found its way into our extant literary sources.

 

6

 

claim (condictio), which was at times stated under oath (sacramentum). Plaintiffs had to resort to one of these remedies (actiones) according to the nature of each claim. If none of these remedies fitted the case, there was no claim. Some remedies were more flexible than others. However, because of its excessive formalism, this archaic system of civil procedure underwent a natural evolution while keeping some of its key features: the two-step procedure, first in front of the magistrate (in iure), then before a judge (apud iudicem); and the turning point of the joinder of issue (litis contestatio), whereby the parties agreed on the legal framework, sanctioned by the magistrate, within which the appointed judge(s) would have to evaluate the facts. This legal framework was eventually described in a written formula instructing the judge(s) about the path to follow and leading to two opposite outcomes of the forthcoming trial (‘if it occurs that ...., then condemn; if not, then absolve’). Whatever happened thereafter, the joinder of issue extinguished the plaintiff’s claim.24

Commercial transactions could often, but not always, be enforced through the existing set of original remedies. Some situations however called for new solutions. Thus, additional remedies came to be created over the next centuries by statutes (e lege), such as the lex Marcia against usury in 104 BC (Gai., Inst. 4.23), or by some magistrate’s edicts. Therefore, remedies were divided between actiones civiles and honorariae (more specifically, actiones aediliciae or praetoriae). As a result, remedies available to parties tended to mushroom, since magistrates were free to deliver – or not – a legal remedy upon one party’s request (petitio actionis). Since the edict was valid for the duration of the magistrate’s tenure of office, an existing actio could technically be denied (denegatio actionis), although social pressure may have played a part in the quest for consistency. Adventurous magistrates would be expected to create new actiones if and when they considered that the situation and/or their sense of equity allowed them to do so. Creativity could take various forms. When a remedy existed (actio directa) but did not exactly match the situation, the magistrate could make adjustment by issuing an actio utilis that extended the scope of the original remedy. If the unsatisfactory existing remedy was an actio civilis (i.e. based on statute), the magistrate could introduce a fictitious element into its intentio to have it fit the situation, such as Roman citizen’s status for one of the parties even though he was an alien (peregrinus). The adjusted remedy would then be regarded as an actio ficticia. If Roman law provided no previously existing remedy to address the issue even remotely, the magistrate could create one based on his perception of the situation (actio in factum). Faced with an actual legal problem, the magistrate was allowed – and expected – to devise a legal solution, leaving it to the judge to decide whether the facts that had led to the solution were correct. As will be seen later on, this instrument proved to be most efficient in dealing with issues related to commercial life.25

It is obvious that such a system, introduced between the fourth and second century BC and called ordo or formulary system, gave great power to magistrates in charge of delivering actiones. To alleviate the suspicion of arbitrariness and to give a sense of coherence to the administration of justice, magistrates with iurisdictio were required to

24 Gaius, Inst. 11-30; and Borkowski – Du Plessis (2005), 63-83.

25 Gai., Inst. 4.30-47; Gaudemet (1982), 615-20.

 

7

 

announce ahead of time when and in what circumstances they would grant a remedy (actio). Magistrates did so through their (yearly) edict. We know next to nothing about the circumstances in which they devised it, but we can imagine that outside inspiration or pressure may not have been totally foreign to their decisions. Whose pressure remains debatable, but jurists and professionals involved in all kinds of business transactions

unavoidably come to mind (e.g. Cic., 2Verr. 1.119).26 Edictal law should be looked at as the ad hoc answer to foreseen or encountered legal problems and situations. In this sense, it can be said that the Roman law of commerce, at least for its private law components, was essentially a law originating with traders and adapted to the requirements of traders, though not exclusively them. The ius honorarium introduced by some magistrates in order to abet, supplement, or correct the existing (civil) law was geared toward protecting and promoting the interests of the community.27 Economic interests must have ranked high on the list. Unfortunately, the details mostly escape us.

During the fifth and early fourth centuries BC, justice was administered by the consuls. Admittedly, plebeian aediles, two in number, may have had some jurisdiction, the extent of which is unclear. In 367 BC, if not before, new magistracies reserved to patricians were created: both the praetorship and the curule aedileship were devised to compensate for the loss of power resulting from sharing the consulship with plebeians. The new magistracies distracted some of the consuls’ powers, thus allowing patricians to retain control over them. As holders of potestas, including the right to issue edicts (ius edicendi) and to enforce their authority (coercitio), and imperium, the basis of their judicial power (iurisdictio),28 these new magistrates, with the addition of more praetors in and after 241 BC who dealt with foreigners (peregrini) and took charge of provincial governorships, were responsible for legal matters and for the supervision of markets. Their part in the development and implementation of the Roman law of commerce is attested by what remains of the edicts they promulgated over several centuries. I would contend that Roman business law is first and foremost edictal law. However, individual edicts were necessarily phrased as briefly as possible and therefore left much room for interpretation. This is where the jurists stepped in, and edictal law must be approached through the juristic writings of the classical period, often the very source from which edictal law can be reconstructed. In the next sections, both sources of law will be jointly examined in order to assess the making and refining of Roman business law. One should remember however that several centuries may have elapsed from the time a remedy was created until the time the legal texts commenting on it were written, before finding their way into the Digest.

Aedilician, praetorian, and gubernatorial edicts

a) edictum aedilium curulium

26 The evidence is clearer for a later period, cf. Nov. Iust. 136 (AD 535); 106 (540) and 110 (541); Ed. Iust. 7 (542) and 9 (date unknown) for the influence of moneylenders on imperial legislation. Cf. Jones (1964), 350 and 1139, n. 63.

27 Papinianus (2 definitionum) D. 1.1.7 and Marcianus (1 inst.) D. 1.1.8.

28 The question of the curule aediles’ imperium is debated, cf. Impallomeni (1955), 109-21; and Reduzzi Merola (2001), 325, n. 18.

 

8

 

In their capacity of city magistrates, the curule aediles were in charge of supervising markets. What remains of their edicts (FIRA I2 66) goes back at least to the first half of the second century BC and is known through a couple of citations preserved in the Digest (D. 21.1, De aedilicio edicto et redhibitione et quanti minoris) and in the work of the second-century

AD antiquarian Aulus Gellius (Noctes Atticae 4.2.1).29 These citations can be divided into two subjects: the sale of slaves (mancipia) and the sale of animals (iumenta, ferae). The edict stipulates a) that buyers of defective slaves will be granted a remedy for rescission (actio redhibitoria) if the seller has concealed a specific defect, whether or not he was aware of it; and b) that a remedy for diminution of the price (actio aestimatoria or quanti minoris) would be available to buyers who had been cheated of what was owed to them, including advertised qualities that turn out to be lacking. A claim of the former kind was valid for sixty days or six months, of the latter for a whole year.30

The aedilician edict was traditional (tralaticium): aediles were in charge during one year only and they usually did not have the legal training necessary to be innovative. Like their better-known senior colleagues, the praetors, aediles had the tendency to borrow most or all of their edicts from their predecessors. This habit resulted with time into a rather static document, eventually codified by Salvius Iulianus in the age of Hadrian, perhaps as an appendix to the praetorian edict (edictum perpetuum). We do not know when the aedilician edict stopped being modified, but some first-century AD documentary evidence suggests that it was still viewed as dynamic under Nero, if not later.31

The content of the aedilician edict is mostly lost, perhaps due to its eventual merging with the praetorian edict. However, the reconstructed text gives several examples of what a defective slave may be. The aediles are said to have done everything possible to avoid ambiguity (Ulp. [1 ad ed. aed. cur.] D. 21.1.1.7), but there was ample room for elaboration. For all its flexibility and pragmatism, edictal law would be inadequate without the jurists’ interpretation. Faced with provisions phrased briefly and with a rather general scope in mind, the jurists took on themselves, over several centuries, the task to adjust these provisions to the requirements of social and economic life, within a logical framework. Title 21.1 of the Digest includes sixty-two excerpts from classical juristic writings on the subject. The earliest authority quoted in them is Cato (presumably the Elder, in [ibid.] D. 21.1.10.1), followed by several Republican jurists. It provides precious evidence about ancient slavery and the slave trade. The late Republican jurist C. Trebatius Testa, for instance, downplayed bad breath as the result of poor oral hygiene (D. 21.1.12.4 [ibid.]), a condition that Apuleius (Apol. 6) could have treated adequately two centuries later. The Flavian jurist Sextus Pedius discusses the case of the bed wetter, and distinguishes between slaves suffering from a bladder condition and those who are too drunk or too lazy to get up at night (D. 21.1.14.4, cited with approval by Ulpian [ibid.]). Mental addictions are taken

29 Impallomeni (1955), 90-136; Pugsley (1974).

30 Gaius (1 ad ed. aed. cur.) D. 21.1.28; Ulp. (1 ad ed. aed. cur.) D. 21.1.19.6 and (2 ad ed. aed. cur.) D. 21.1.38 pr. (about iumenta) .

31 Reduzzi-Merola (2001), based on TPSulp. 43 (21 Aug. 38) (possibly also TPSulp 42 and 44); TH 59-62 (between AD 47 and 63); and Petronius, Sat. 53.9-10. Cf. de la Hoz Montoya (2008), 376-80.

 

9

 

seriously: slaves hooked on games or art work are considered defective by the late second-century AD jurist Venuleius Saturninus ([5 act.] D. 21.1.65 pr.). At stake were the smoothness of the slave trade, consumer protection, and the expected productivity of slave labor.

The animal trade presented similar problems, but seemingly triggered less discussion on the part of the jurists (D. 21.1.38, 40, and 41): dangerous animals, including dogs, should be kept away or chained, so that they do not attack people, thus causing damages calling for monetary penalties. Animals should be sold with the trappings (ornamenta) they wore at the time of sale, lest the sale be rescinded or a diminution of the price be granted. Interestingly, the aediles consider sales (and returns) in bulk, for instance for a pair of mule. This is valid for slaves as well, be it a company of actors or, merely, siblings (Ulp. [ibid.] D. 21.1.38.1 and 14; Paul [1 ad ed. aed. cur.] D. 21.1.39).

Diminution of the price could be obtained through the actio quanti minoris/aestimatoria, which the aediles mentioned only in connection with animals (D. 21.1.38 pr. and 13), but which the jurists extended to slaves.32 This is typical of the work of classical jurists. We know of provisions dealing with companies of slave dealers (societates venaliciariorum, Paul [2 ad ed. aed. cur.] D. 21.1.44.1), the castration of young slaves (Ulp. [18 ad ed.] D. 9.2.27.28), or the definition of trappings (ornamenta, Paul [2 ad ed. aed. cur.] D. 50.16.74). Consumer protection, however, was not limited to the sale of slaves or animals (at least by Diocletian’s time, CI. 4.58.4 [286-293]). Besides, Ulpian points out that the aedilician edict applies to sales only, not to other categories of contracts such as hire and lease (locatio conductio), because such contracts were never under aedilician jurisdiction or – and the introduction of an alternate explanation would be telling if the whole passage were not interpolated – because both contracts are different (‘non similiter,’ Ulp. [1 ad ed. aed. cur.] D. 21.1.63).

Even though aediles were traditionally in charge of supervising local markets and their edicts were recognized as valid not only in Rome, but also in the rest of Italy and apparently across the Empire by the second century AD,33 the scope of aedilician law was dwarfed by praetorian law. This is confirmed by the fact that only few classical jurists are known to have commented on the aedilician edict: Ofilius and possibly Labeo in the late Republican and Augustan periods, Caelius Sabinus (cos. 69) and Sextus Pedius in the first century AD, Pomponius and Gaius in the second, Paulus and Ulpian in the early third. It is likely that the aedilician edict eventually became an appendix to the praetorian edict, since

32 Gaius (1 ad ed. aed. cur.) D. 21.1.18 (false advertisement); Ulp. (1 ad ed. aed. cur.) 21.1.19.6 (respective deadlines, six months for a. redhibitoria, one year for a. quanti minoris); (ibid.) 21.1.31.5 (citing Pomponiius, collective purchase), 10 (idem), and 16 (successive claims by buyer); Pomponius (23 ad Sab.) 21.1.36 (about bulk price); Paul (1 ad ed. aed. cur.) 21.1.43.6 (about the complementariness of both remedies); Paul (11 ad Sab.) 21.1.47 pr. (extinction of the buyer’s claim after the slave’s manumission); and Ulp. (80 ad ed.) 21.1.61 (for an undeclared servitude).

33 Cf. above n. 27, and FIRA III2 87-88 (AD 139 et 142, Dacia); FIRA III2 133 (AD 151, AD 151) and 132 (Seleucia in Pieria, AD 166).

 

10

 

the curule aediles’ competences were progressively absorbed by other magistrates and imperial officials.34

b) edictum praetoris

Both urban and peregrine praetors issued edicts, general rules and specific remedies, that were valid and possibly binding during their – one-year – term in office, at the end of which individual edicts could be either dropped or renewed by transfer into the successor’s edict.35 Dio Cassius reports that in 67 BC a plebeian tribune named C. Cornelius introduced a plebiscite (Lex Cornelia de edictis/de iurisdictione praetoris) compelling all praetors – possibly governors as well – to abide by their own edicts, which contained the basic principles (dikaia) according to which they would administer justice, but not all remedies (dikaiomata) which were required in order to enforce contracts. This piece of legislation was part of a larger package aimed at curbing corruption on the part of the senatorial class and ensuring legal consistency. It may have reenacted an earlier custom or law that had been neglected or broken during the 80s and 70s BC.36

The making of praetorian edicts during the Republic and early Empire is somewhat of

a mystery.37 Of all preserved or reconstructed edicts, none can be dated precisely, and few only roughly. Vague termini ante quos are provided by quotations, mentions, or allusions in the commentaries to the edict by Republican and Augustan jurists, such as Servius Sulpicius Rufus, Ofilius, or Labeo. The sum of those edicts that had been made permanent over the years was codified ca. AD 130 into the Edictum Perpetuum, again by Salvius Iulianus. What remains of it (FIRA I2 65) is known as the edict of the urban praetor (EP) and has been reconstructed by O. Lenel in the nineteenth century on the basis of the organization of Justinian’s Code and Digest and the numerous quotations preserved in the latter, especially excerpted from large commentaries by Paul (80 books), and Ulpian (81 books). What is left is tantamount to 292 entries (rubricae) distributed in 45 titles (tituli) and five parts (partes). For many of the 292 entries, nothing but the title is preserved or can be reconstructed. The listing is not necessarily representative of the relative chronology of their introduction into the edict. Actually, there are reasons to believe that the arrangement was revisited even shortly before, if not at the time of, its final codification. On the basis of a comparison of the space devoted to various parts of the Edictum Perpetuum in Paul’s and Ulpian’s commentaries as opposed to Sextus Pedius’ commentary on the first-century-AD edict, it appears that some dispositions regarding business law may have been shifted to a different section of the edict. Thus, sometimes in the late first or early second century AD, possibly at the time of the codification of the Edictum Perpetuum by Salvius Iulianus, the law of

34 Dio 53.2.2 and 54.2.3. Cf. Giachi (2005), 65-70, esp. 67, n. 147.

35 Brennan 2000. On the peregrine praetor, cf. Daube (1951); and Serrao (1954).

36 Dio Cassius 36.40.1-2; Asconius, Ad Cic. Pro Cornelio de maiestate 59.8-9 (Clark) = 48 (Stangl); and Cic., 2Verr. 1.46.119 about Verres deciding against his own edict. Cf. Rotondi (1912/1990), 371; Griffin (1973), 209; Pinna Parpaglia (1987) and (1992); and Palazzolo (1991).

37 Guarino (1980), esp. 68-76 for the distinction between edictum perpetuum (promulgated at the beginning to the year of office) and edictum repentinum (promulgated during the year of office), as opposed to denegatio actionis (denial of a remedy); and Mantovani (2000).

11

 

indirect agency (cf. below) seems to have been severed from its original context, i.e. the special liability of seamen, innkeepers, and stablekeepers (= managers of relay-stations) for what has been entrusted to them in connection with the practice of their trade, to be eventually linked with banking and financing on the one hand, and with consensual (good faith) contracts on the other.38

The urban praetor’s edict contributes a lot to our knowledge of Roman business law, as nine titles out of forty-five, and more than forty entries out of 292, deal with legal issues concerning commerce.39 Unsurprisingly, this means maritime law, banking, agency, contracts among private individuals, and between private individuals or companies and the Roman government, securities, and procedure. The question is whether any of the edicts concern traders to the exclusion of other actors participating in legal transactions. The answer is unambiguously positive, although such cases are rather limited.

EP 2.19.106-112 introduced good-faith remedies (De bonae fidei iudiciis) that are central to economic activities and will provide a starting point. Deposit (D. 16.3) as a real contract, trust (Gai., Inst. 2.60) or fiduciary agreement as a form of real security, mandate (D. 17.1), partnership (D. 17.2), sale (D. 19.1), and hire (19.2) as consensual contracts, were all devised during the mid-Republican period, and became standard legal instruments. They advantageously supplemented the older, flexible though formal contract by (verbal) stipulatio in that they considered the will (voluntas), permission (patientia), awareness (scientia), or lack of it (ignorantia) of the parties to a contract. By allowing transactions to be carried out despite the absence of one or both parties, they certainly opened the door to a major innovation in Roman business law, the law of indirect agency (cf. below). The consensual contract of sale (emptio venditio) was introduced by the second century BC. At the minimum, buyer and seller had to reach a specific agreement on both the object of sale and its price. The usual terms of the contract could be modified or specified through stipulatio or pactum. In practice, the parties’ agreement was often made explicit – and strengthened – by the written record (chirographum) of what sounds like a stipulatio. The seller had to warrant for his title to the object of sale and for its quality (i.e. lack of defects). The buyer, who originally bore all the risks of the transaction (“caveat emptor”), came to enjoy the protection of the law, enforced by aediles, praetors, and later on prefects, in accordance with established standards of good faith.40 The Roman law of sale, for all its sophistication and prominent importance in commercial life, was not specific to the business

38 Giachi (2001), 14, n. 23, about the transfer of EP 2.18.101-105 (Quod cum magistro navis, institore eove qui in aliena potestate est, negotium gestum erit) from its hypothetic, original place just after EP 2.15.78 (De his quae cuiusque in bonis sunt – In factum adversus nautas caupones stabularios) toward a place between EP 2.17.95-100 (De rebus creditis) and EP 2.19.106-112 (De bonae fidei iudiciis).

39 EP 1.3.9 (De edendo – argentariae mensae exercitores); EP 1.11.49-50 (De receptis); EP 2.15.78 (De his quae cuiusque in bonis sunt – In factum adversus nautas caupones stabularios); EP 2.17.95-100 (De rebus creditis); EP 2.18.101-105 (Quod cum magistro navis, institore eove qui in aliena potestate est, negotium gestum erit); EP 2.19.106-112 (De bonae fidei iudiciis); EP 3.32.183-185 (De publicanis); EP 4.40.218-223 (Quemadmodum a bonorum emptore vel contra eum agatur); EP 5.44.269-279 (De exceptionibus).

40 TPSulp. 42 (Puteoli, AD 26); 43 (Puteoli, AD 38); and 44 (Volturnum, first c. AD); and FIRA III2 132-142 (all documents of early and late imperial date). On sale, cf. Gai., Inst. 3.139-141; and D. 18.1-19.1; Crook (1967), 215-21; Johnston (1999), 79-84; and Zimmermann (1990), 230-337.

 

12

 

community.41 Other areas of law, such as transportation or agency are more likely to have been the preserve of merchants.

The law of carriage, for both land- or sea-transport of people and goods, combines several areas of law, including the consensual contract of hire (locatio conductio, D. 19.2; C. 4.65), whose object could be either the means of transportation (res in the form of pack- or draft-animals, mounts, porters, wagons, ships, storage room), the task (opus) or the services provided by professionals (operae); specific modes of financing (maritime loans or faenus nauticum, D. 22.2; CI. 4.33) combined with early forms of insurance against the enormous risks connected with navigation (lex Rhodia de iactu, D. 14.2) (cf. below); and safekeeping (custodia) in storage (horrea) or in transit (naves, cauponiae, stabula, D. 4.9 and 47.5). EP 1.11.49 (De receptis) and 2.15.78 (In factum adversus nautas caupones stabularios) call for a higher standard of liability on the part of shippers and inn-/stablekeepers for the goods entrusted to their care in the context of their professional activities. The operators (exercitores) of the ship, inn, or stable are responsible for the wrongdoings (delicta, including theft/furtum) of their employees, slaves or free, because they chose their staff and had a chance to vet them. By analogy, innkeepers are also responsible for the wrongdoings of their guests, at least if they live there on a permanent basis, unlike passing travelers and passengers on a ship (Ulp. [38 ad ed.] D. 47.5.1.6).42 The owner of the stolen good(s) can sue either the thief iure civili or the operator honorario iure (ibid. D. 47.5.1.3). If however the operator has guaranteed his safekeeping of the goods, he can sue the thief himself (D. 47.5.1.4).

The duty of safekeeping was conditioned by the operators’ free choice (arbitrium) in accepting the goods to be watched. This is what distinguishes traders from thieves: the former, unlike the latter, must show good faith (fides) and refrain from fraud. In that sense, the praetor’s edict is considered “most useful” by Ulpian ([14 ad ed.] D. 4.9.1.1). The phrasing is however most economical, and the terminology is to be elucidated by jurists: ‘seaman’ (nauta) designates the operator of a ship (exercitor) or his agent (magister navis), not the crew. Along the same line, the ‘inn-’ or ‘stablekeeper’ (caupo and stabularius) is the person in charge of the facilities, either as operator (exercitor) or as manager (institor) (ibid. D. 4.9.1.2-3 and 5); ‘ship’ is understood as sealiner, freighter, river-boat, or raft (Labeo, cited ibid. D. 4.9.1.4); ‘goods’ (res, merces) mean not only merchandises (mercedes), but also personal belongings transported as luggage or clothing (Vivianus and Pomponius, ibid. D. 4.9.1.6-8); etc. The praetor provides a general ruling, the jurists determine its scope, and the judge sees to its application.

In principle, the remedy brought against the operator/manager is based on the initial contract of hire and lease or deposit that binds him to the plaintiff, and consequently calls for different standards of liability, fault (culpa) in the former case, fraud (dolus) in the latter (Ulp. [14 ad ed.] D. 4.9.3.1). Cases not covered by the edict, for instance for lack of fault or fraud on the part of the operator or because no price was paid for the service, gave rise to

41 With the possible exception of the sale of wine, cf. Frier (1983), 292.

42 A. Petrucci kindly pointed out to me that Gaius ([5 ad ed. prov.] D. 4.9.2) includes the innkeeper’s liability for theft committed by travelers (viatores) in case of receptum. Ulpian follows suit for passengers (vectores) in similar condition ([14 ad ed.] D. 4.9.1.8).

 

13

 

an actio in factum (ibid. and Paul [22 ad ed.] D. 4.9.6). If the damage was caused by one of the sailors’ slave who was not a sailor himself, the operator would nevertheless be liable to an actio utilis (Ulp. [18 ad ed.] D. 4.9.7.3). Conversely, operators could avail themselves of special legal protection (exceptio) in case of an act of God (vis maior), such as shipwreck or pirate/bandit attack (Labeo, cited by Ulp. [14 ad ed.] D. 4.9.3.1). Thus the law took into account the reality of commercial life by striking a balance between the interests of customers and those of business people.43

The praetorian edict includes two dispositions concerning professional bankers (argentarii, argentariae mensae exercitores). The so-called receptum argentarii/argentariorum (EP 1.11.50) was, like the receptum nautarum cauponum stabulariorum discussed above, an informal promise or guarantee (pactum praetorium), in this case, to pay a client’s debt on an agreed day. Surprisingly, this arrangement binds banker and creditor, and leaves out the customer/debtor. In addition, the nature and very existence of the debt are irrelevant. Despite its obvious usefulness for banking, the receptum argentariorum is little attested in the legal sources, because its was merged in late Antiquity with the wider ranging, less specific constitutum debiti.44

The money that the banker agrees to pay probably comes from the customer’s account. The praetor rightly compels bankers (argentarii, argentariae mensae exercitores) to produce accounts upon request from judicial authorities (EP 1.3.9a and b, De edendo). The reasoning behind the edict is considered most fair by Ulpian ([4 ad ed.] D. 2.13.4 pr.-1). The praetor arranged for reciprocity, the banker being entitled to ask for the production of accounts on the part of an opponent, unless the banker had the means to achieve the same result on the basis of documents readily accessible to him because of his occupation (ibid.

D. 2.13.6.8-9).45 Gaius, in the first book of his commentary to the provincial edict (D. 2.13.10.1), explains that bankers (argentarii) have a special obligation to produce their accounts because their trade has a publica causa, which means that the Roman people had a vested interest in regulating the profession.46 For that reason women were banned from it

according to the third-century jurist Callistratus.47

EP 2.17.95-100 (De rebus creditis) deal with various forms of loans, for consumption (mutuum, D. 12.1; CI. 4.1-2) or for use (commodatum, D. 13.6; 4.23), pledge (pignus, D. 13.7;

43 Huvelin (1929), 115-59; Zimmermann (1990), 514-26 (receptum nautarum cauponum stabulariorum); Gröschler (2002), 70-79.

44 D. 13.5 and CI. 4.18 (De pecunia constituta). Cf. Crook (1967), 232-33 and 243; Zimmermann (1990), 511-14; Andreau (1987), 597-602; id. (1999), 43-44 and 58, with reference to TP 151 (= FIRA III2 131e = CIL IV 3340.91, AD 62) as a possible specimen from legal practice; Petrucci (1991), 378-83; id. (2002), 57-65; and id., in Cerami et al. (2004), 129-35.

45 EP 1.3.9 (De edendo – argentariae mensae exercitores); cf. Andreau (1987), 551 and passim; Andreau (1999), 30-49, esp. 46; Petrucci (1991), 141-71; id. (2002), 23-27 and 140-53; id., in Cerami et al. (2004), 175-86.

46 Petrucci (2002), 18 and 123, n. 30. On causa as a ground for legal action, cf. Borkowski – du Plessis (2005), 258-59, with reference to Ulp. (4 ad ed.) D. 2.14.7.4.

47 ‘Opera virilis’: Callistratus (1 edicti monitorii) D. 2.13.12. Cf. Andreau (1987), 497. The nature of the edictum monitorium is uncertain.

 

14

 

CI. 4.24), and setoff (compensatio, D. 16.2; CI. 4.31). There is no questioning the usefulness of these legal institutions for commercial life, in terms of credit, security, and payment, but were by no means restricted to the activities of professional traders. It is however tempting to explain the sophistication and sometimes paradoxical and adventurous nature of some arrangement as dictated by the requirement of specialized trading.48 It is not entirely surprising that deposit (depositum) is dealt with in what could have been the very next entry of the edict (EP 2.19.106), before the shift of the title dealing with agency assumedly occurred.49

Agency is undoubtedly one of the most significant areas of progress in the field of Roman business law.50 As a matter of principle Roman law was adverse to the idea that a person’s action could engage another’s liability, while there was no problem with benefiting from it. This somewhat contradictory position precludes the concept of agency, so important in economic life, even more so given the negative attitude of the Roman elite towards trade and commerce (Cic., Off. 1.150-151). The patriarchal nature of Roman society offered a way to overcome this obstacle: persons in power (in potestate, alieni iuris), such as sons, daughters, other descendants, and mostly slaves had no legal capacity of their own, and therefore participated in that of their pater familias/dominus. Slaves’ wrongdoings (delicta) originally gave rise to noxal liability on the part of their master.51 The praetor, aware of this asymetrical state of affairs, had only to extend the liability of the principal to the contracts of his agent.52 To that effect a set of six remedies was created over the course of time, presumably between the late third and early first centuries BC: EP 2.18.100¬105 (Quod cum magistro navis institore eove qui in aliena potestate est negotium gestum erit - On the dealings of the ship’s captain, business manager, and person in power) and contains one of the boldest and most ingenious creations of praetorian law,53 the early history of which is unfortunately blurred, but can be reconstructed on the basis of a few classical – therefore much later – legal texts.

Ulpian ([29 ad ed.], D. 15.1.1 pr.) records that the praetor attended first to remedies given for the full amount (in solidum) on the basis of contracts concluded with persons in power. Remedies giving rise to a limited liability (dumtaxat de peculio aut de in rem verso) and those based on the contracts of non-dependent persons (sui iuris, extranei) should therefore be regarded as later additions or extensions. Gaius (Inst. 4.70-71) explicitly says that the praetor started (‘inprimis’) with the grant of a remedy on authorized transactions carried by a dependent (actio quod iussu) and added by analogy (‘eadem ratione’) two further remedies

48 Sirks (forthcoming), with reference to Ulp. (26 ad ed.) D. 12.1.9.8; Celsus (5 dig.) D. 12.1.32; and Ulp. (30 ad ed.) D. 16.3.7.2 (about the insolvency of nummularii).

49 Cf. above, with Giachi (2001).

50 Serrao (1989); Zimmermann (1990), 34-67, esp. 45-58; Aubert (1994); Cerami, in Cerami et al. (2010), 34-63.

51 Cf. above, XII Tables 8.2 and 12.2 (Crawford); Ulp. (14 ad ed.) D. 4.9.3.3; and id. (18 ad ed.) D. 4.9.7.4.

52 Ulp. (28 ad ed.) D. 14.3.1. The blurring of criminal and civil liability is touched upon by Ulp. (1 ad ed. aed. cur.) D. 21.1.23.4-5.

53 The latest work on this part of the edict (formulae) is Micelli (2001), esp. 185-228.

 

15

 

for the transactions concluded by dependent ship’s captains (actio exercitoria) or business managers (actio institoria). The liability of the father/master was based on his willingness (voluntas) to allow contracts to be made with his dependent. This willingness was expressed through the appointment (praepositio) of the ship’s captain (magister navis) to a ship (navis) and of the business manager (institor) to a business (taberna aut quaelibet alia negotiatio). It is possible though not sure that originally the actio institoria was available only in the context of

the management of a shop (taberna).54 Alternatively, a rural context, such as the Catonian villa, may have provided the original Sitz im Leben of the actio institoria, since non-legal sources refer to institores as vilici and actores.55 The extension to other negotiationes, such as workshops (officinae, fabricae), credit institutions (mensae), etc., would only be natural.

Whether ships should be regarded as a later extension is a controversial question. Gaius (Inst. 4.71) and Justinian’s Digest (D. 14.1 and 3) and Code (CI. 4.25) invariably present the actio exercitoria before the actio institoria. This suggests that by the time of the composition of the Edictum Perpetuum (ca. AD 130) the order of presentation reflected the prominence – not necessarily the priority – of the actio exercitoria over the actio institoria. On the other hand, both Gaius (4.71) and Ulpian ([28 ad ed.] D. 14.1.1 pr.) stress the specificity of the conditions in which a ship’s captain works in comparison with a business manager: the distance separating agent and principal from one another in the context of seaborne trade makes it more difficult for third parties to check the agent’s legal status and sphere of competence. Subappointments are more readily acceptable in the case of ship’s captains than business managers, ‘for practical reasons’ (‘propter utilitatem’, ibid. D. 14.1.1.5). It also seems that, by contrast with business managers, ship’s captains are presented as non-dependents, which I believe to be the result of a later development: third parties could elect whether to sue the ship’s operator (exercitor) or the ship’s captain; the operator had no remedy against third parties contracting with his captain, because he allegedly did not need any, since he could sue the captain on the contract of employment or mandate which defined the relationship between principal and agent. Ulpian notes however that ‘in practice’ (‘plane’) the prefects in charge of the corn supply (annona) and provincial governors helped them in an informal way.56 The specificity of agency in the context of seaborne trade explains why two distinct but related remedies were necessary, and may suggest that the actio exercitoria developed from a more general actio institoria into a ground¬breaking legal instrument, both remedies eventually applying to the contracts of non¬dependent agents.57

The main legal issue discussed by the jurists in connection with both remedies concerns the scope of the appointment (praepositio). In order to give rise to the principal’s full liability, the contract concluded by the agent must pertain to the business he or she –

54 Földi (2001), 78-84.

55 Aubert (1994), 117-200.

56 Ulp. (28 ad ed.) D. 14.1.1.17-18 (‘extra ordinem’) to be compared with Ulp. (28 ad ed.) D. 14.3.1, citing the late second-century jurist Marcellus, and Gaius (9 ad ed. prov.) D. 14.3.2. Cf. Sirks (2002), 139.

57 Aubert (1999a). On the actio exercitoria, cf. Gaurier (2004), 79-95.

 

16

 

women and children could be appointed58 – is in charge of. Republican and Augustan jurists, such as Servius and Labeo, and their successors list various types of facilities in connection with which specific transactions may give rise to an actio institoria/exercitoria, thus contributing to the definition of business and enterprise in the Roman world.59 The scope of the appointment was implicitly conditioned by the nature of the business, and could be explicitly spelled out in a charter (lex praepositionis) used as a job description.60 Any extension (iussum) or limitation (proscriptio) had to be publicly posted and advertised. Double appointments – combining two (or more) types of activities under the responsibility of the same manager – are attested.61

Ulpian also discusses cases of subappointment ([28 ad ed.] D. 14.1.1.5), of joint managers working for a single operator (ibid., D. 14.1.1.13-14 and 14.3.11.5), and of joint operators appointing a single manager, either within a partnership (societas) or as joint owners (ibid., D. 14.3.13.2). The latter configuration was identified as the possible ancestor of company law, so prominent in modern commercial/business law.62 The Roman law of indirect agency also applied to collectivities, both private and public, such as companies of publicans (societates publicanorum), professional and religious associations (collegia), and towns (municipia and coloniae).63

The second part of the praetorian edict on indirect agency deals with those remedies that strictly apply to transactions carried out by dependents. The edictum triplex (Ulp. [29 ad ed.] D. 15.1.1.1) includes the actio quod iussu (D. 15.4), mentioned above as the likely original remedy for full liability of the principal, and two additional, eventually intertwined remedies (actio de peculio aut de in rem verso, D. 15.1-3), which called for a limited liability on the part of the principal on account of the contracts concluded by and with his dependent.64 Connected with the actio de peculio was a sixth remedy (actio tributoria, D. 14.4), whereby the principal was treated like any of the creditors when the agent’s insolvency gave rise to an

action on his peculium.65 The sheer order of presentation of the various remedies in the Digest (D. 14 and 15) and assumedly in the Edictum Perpetuum (EP 2.18.101-105) shows that

58 Ulp. (28 ad ed.) D. 14.1.1.21 and 14.3.7.1; Gaius (9 ad prov. ed.) D. 14.3.8. Cf. Aubert (1994), 43, 56, 140-41, 193, 372, 224-26, 292-93, and 419-20. The impact of the SC Velleianum (ca. AD 46) (EP 2.18.105; D. 16.1; CI: 4.29) on the ability of women to act as agents is unclear to me.

59 Ulp. (28 ad ed.) D. 14.3.5.1-15 and 14.3.13 pr.; Paul (29 ad ed.) D. 14.3.16 and id. (30 ad ed.) D. 14.3.17 pr., etc.

60 For lex praepositionis as a kind of lex contractus, cf. Aubert (2003); id. (2005a); id. (2005b); and id. (2007a); Aubert – Raepsaet (2011).

61 Ulp. (28 ad ed.) D. 14.1.1.12 (certa lex); ibid. D. 14.3.11.2-6 (proscriptio); ibid. 14.3.13 pr. (double appointment); and Gaius (9 ad ed. prov.) D. 14.5.1 (iussum). Cf. Aubert (1993); id. (1994), 6-14, 50-52, and 335; id. (2003); id. (2005a); id. (2005b); id. (2007a); Jakab (2008); and Aubert – Raepsaet (2011) (mandate).

62 Di Porto (1984), 169-204; Aubert (1994), 54-57 and 62-63, with additional references, to which Paul (4 ad Plaut.) D. 14.3.14 should be added. For Roman company law, cf. Crook (1967), 229-36.

63 Aubert (1994), 325-47; id. (1999b). On publicans, cf. EP 3.32.183-185; D. 39.4, with Gaius’ commentary ad edictum praetoris titulo de publicanis; and Maganzani (2002).

64 Actio quod iussu: Schleppinghoff (1996); and Coppola Bisazza (2003). Actio de peculio: Aubert (2010), with earlier bibliography. Actio de in rem verso: Chiusi (2001).

65 Actio tributoria: Chiusi (1993).

17

 

it cannot reflect the chronology of their respective creation. The transfer of the actia quad iussu to the very end of the series (D. 15.4) suggests that the order of presentation reflects the relative importance of each individual remedy in comparison with one another, the actia quad iussu being at best subsidiary in the classical period. Therefore it is telling that both commercial remedies (actianes exercitaria and institaria) applying to the contracts concluded by both dependents and non-dependents had taken precedence over all the others by the time of redaction of the Edictum Perpetuum in the second century AD. Accessorily, the reconstructed relative chronology of the creation of the so-called actianes adiecticiae qualitatis (with the addition of the actia tributaria) and their respective order in the Edictum Perpetuum indicate that the praetor favored the interests of third parties contracting with dependent agents over those of principals whose liability was engaged by their dependents’ transactions.

Gaius, in the ninth book of his commentary on the provincial edict, to which we will move shortly, reports that governors (pracansules) ensured that those who contracted with persons in power obtain their due: if the commercial remedies (a. exercitaria, institaria, and, paradoxically, tributaria) did not apply, the governor would grant a remedy for full liability on authorized transaction (a. quad iussu), provided such authorization existed, or on the enrichment of the family estate (a. de in rem verso), or, in the last resort, on the peculium (D. 14.5.1). The order of preference is altogether clear, and reflects the sense of equity (‘ex bono et aequo’) of both the third party and the political/judicial authority. In a separate case, the master of a runaway slave appointed to lend money and to accept security (pignus) fought off a suit from barley traders who had been promised payment on behalf of customers. Interestingly, the prefect of the corn supply stepped in on behalf of the traders and decided to hold the master liable in full, pointing out that the slave was notoriously in the habit of being involved in various businesses, such as renting warehouses (harrea). The early-third-century jurist Paul, who reports the case in which he seemed to have acted as the master’s lawyer, unsuccessfully argued for considering the slave’s payment as a type of guarantee (fideiussia), as opposed, presumably, to a receptum argentarii (cf. above). On appeal, the prefect’s decision was upheld by the emperor (Paul [1 decretarum] D. 14.5.8).

c) edictum pravinciale

Ancient legal commentaries leave no doubt that the bulk of edictal law pertaining to commercial life was established through the praetorian edict, before being developed by the jurists. As the case discussed in the last paragraph shows, provincial governors and prefects also used their ius edicendi to intervene when needed. Gaius’ lone commentary on the provincial edict (in thirty books vs. only ten for his commentary on the praetorian edict)66 suggests that a general provincial edict, distinct from individual edicts applying to distinct provinces, existed by the second century, but that it was not so different from the praetorian edict to justify a separate treatment by any other classical jurist. Like the edict of the curule aediles, it may have been absorbed in the Edictum Perpetuum at one point.

66 Martini (1969), 103-28; Santaluccia (1975).

18

 

The situation may have been different in the Republican period. We happen to know a bit about Cicero’s own edict as governor of Cilicia in 51/50 BC through his letters to his friend Atticus. In the winter of 50, Cicero was alternatively approached by the Salaminians of Cyprus and by M. Scaptius and P. Matinius, M. Iunius Brutus’ agents, for the recovery of a debt owed by the former to the latter. The disagreement bore on the interest to be paid. Cicero argued that he had promised (in his edictum translaticium) that he would not allow for more than twelve percent yearly compound interest (Att. 5.21.10: ‘cum anatocismo anniversario’), whereas Scaptius was asking for 48% simple interest. Scaptius opposed to Cicero’s edict a senatorial decree passed a few years before (56 BC) compelling the governor of Cilicia to honor the bond (syngrapha) as such, in blatant contradiction with the lex Gabinia of 68/67 BC forbidding Romans to lend money to provincial communities. The details of the story and its outcome should not concern us here. Cicero obviously felt constrained by the terms of his own edict (described in a later letter, Att. 6.1.15-16), though not by those of his predecessor in the position of governor of Cilicia. Cicero’s provincial edict, kept intentionally short, was derived from the Asiatic edict of Q. Mucius Scaevola (cos. 86 BC) and divided into two parts (genera): one was considered exclusively provincial and dealt with civic accounts (rationes civitatum), debt (aes alienum), rate of interest (usura), contracts (syngraphae), and regulations applying to publicans; the other part, dealing with inheritance, possession and sale of goods, etc., contained usual edictal material. Some rules remained unwritten, but Cicero boasts that he would let provincials use their own laws in their own courts, thus maintaining the fiction of restored autonomy. In disputes between publicans and provincials, the governor’s protection of the latter amounted to no more than a temporary measure, advertised as such to pressure the Greeks to settle as quickly as possible: after a fixed deadline the provision of the provincial edict regarding the rate of interest would give way to the terms of the agreements (pactiones). Provincial governors, like aediles and praetors, used their ius edicendi both before and during their term of office, a double practice allowing them to react to unexpected circumstances and problems while exposing them to undue pressure on the part of groups and individuals, and to the risk of self-contradiction.67 Accessorily, local customs (mos regionis) and laws could be taken into account (Ulp. [24 ad ed.] D. 25.4.1.15, admittedly a different context).

The combined creativity of curule aediles, urban and peregrine praetors, and provincial governors provided ample material for the jurists to work on to adjust the law to the needs of the business community in a growing and increasingly interconnected Mediterranean world. The flexibility of law-making through temporary or permanent edicts combined with a constant, recurring, and diversified exposure to neighboring legal systems, especially in the Greek East, resulted in the blending of the old Roman legal system. Cicero’s Salaminian affair brought him into contact with the Greek institution of the syngrapha, which, like the better attested chirographum, eventually sifted into Roman law as a form of written contract in addition to real, oral, and consensual contracts.68

67 Cic., Att. 5.21.10-13 (13 Febr. 50); 6.1.15-16 (22 Febr. 50); and 6.2.7-10 (early May 50). Rotondi (1912/1990), 373-74 (lex Gabinia). Cf. Pugliese (1964/1985); Martini (1969), 11-102; Peppe (1991); and Maganzani (2007).

68 Gaius, Inst. 3.134; Meyer (2004), 12-19 and 125-68.

 

19

 

Legal borrowing

Unsurprisingly, the Greeks were mostly influential in the field of maritime law. In spite of what the Romano-Carthaginian treaties of the early Republican period (cf. above) may suggest, the Romans were late – in relation to the Greeks and Carthaginians – in developing seaborne trade and their own sea power. The Mediterranean world, especially in the East, was already bursting with commercial activities in the classical and Hellenistic periods. Navigation was regarded as – and to some extent was – a dangerous activity, and the Greeks had devised some legal institutions aimed at minimizing the financial risks attached to it: bottomry loan (foenus nauticum) and compensation (contributio) for jettison (iactus) are two forms of maritime insurance which the Roman jurists elaborated on the basis of allegedly preexisting arrangements familiar to the trading community. How these arrangements passed into Roman law – provided they were actually borrowed and not simply reinvented as the logical solution to a universal problem – is a mystery, but it is fairly clear that the edict was not the way.69

Maritime loans are attested in the Roman world as early as the second century BC.70 The loan (pecunia traiecticia) was made out by private investors or professional bankers – acting as middlemen – to a shipper or group of shippers in a partnership, and had to be paid back only if and when ship and cargo reached the agreed destination. Ship and cargo were considered securities for the loan, whose duration was limited in time (e.g., 200 days in Callimachus’ case, for a trip from Beirut to Brindisi).71 In case of shipwreck, pirate attack, or other acts of God, the creditor would bear the loss. On the other hand, if the navigation was successful, the creditor was entitled to collect interests at a much higher rate than the law permitted for regular loans (12% yearly, cf. above). Many points remain problematic, and unfortunately neither the nine excerpts in the Digest (D. 22.2, from Servius to Ulpian) nor the four Diocletianic constitutions from the Code (CI. 4.33) provide more than hints. It is not clear, for instance, on which ground the legal limit of the rate of interest could be exceeded. We know of no statute, senatorial decree, plebiscite, or edict lifting it, although no less than four laws on usury have been proposed or passed between 217 and 192 BC.72 The jurists and the drafters of imperial constitutions insist on the notion of periculum creditoris (moneylender’s risk), thus revealing a favorable bias toward shippers, who had to show nothing but good faith.73 Maritime loans sound like a necessary evil akin to both speculation and insurance against disasters.

69 Crook (1967), 223-25. Cf. in general Huvelin (1929), 184-218; Rougé (1966); Gaurier (2004), 97-133. On the Greek daneion nautikon, cf. Schuster (2005).

70 Plut., Cato Maior 21.6. Cf. D. 22.2; and CI. 4.33, with Litewski (1973); Biscardi (1974); Casson (1986); id. (1990); Zimmermann (1990), 181-86; and Ankum (2000). This type of loan is occasionally represented in the papyri, cf. for instance SB III 7169 (BL, second cent. BC) for a trip to the Somali coast (with Roman names: ll. 12, 19, 21!); SB III 7170 (id.); SB XIV 11850 (BL, Theadelphia? Febr. 13, AD 149); and SB XVIII 13167 (BL, so-called Muziris papyrus, mid-second cent. AD) for a trip from Alexandria to India and back.

71 Scaevola (28 dig.) D. 45.1.122.1; Johnston (1999), 95-96; and Sirks (2002), 142-49.

72 Rotondi (1912/1990), 99.

73 Moneylenders had a say in legislation in a later period, cf. Nov. Iust. 106 (540) and 110 (541); cf. Jones (1964), 350.

20

 

For disasters did strike, though unevenly. Greek and Roman novels, including Paul’s report in the Acts of the Apostle (27:14-20), betray both fascination for and familiarity with standard procedure during storms on sea. To rescue the ship, part or all of the cargo sometimes had to be jettisoned. Those whose goods had been saved thanks to the others’ – sometimes unwilling – sacrifice were called to contribute to the loss. The Roman jurists acknowledge the routine resort to some foreign usage,74 the so-called lex Rhodia de iactu, the Rhodian origin of which is little more than hypothetical. From what can be reconstructed on the basis of D. 14.2 and Pseudo-Paul’s Sententiae (2.7), it seems that a contribution was expected not only in case of jettison, but also for ransoming from pirates, according to late Republican and Augustan jurists like Servius, Ofilius, and Labeo (Paul [34 ad ed.] D. 14.2.2.3). Those and later classical jurists devised a very sophisticated system of calculating the estimated, respective market value of the goods, either lost, damaged, or saved. Even though the lex Rhodia was not included in any edict and should be regarded as a mere appendix to the law of hire and lease (locatio conductio), the title dedicated to it was placed between the actio exercitoria (D. 14.1) and the actio institoria (D. 14.3) by the compilers of the Digest, because the owner(s) of the jettisoned goods (vectores) would have a remedy against the ship’s captain, and because the lex Rhodia was akin to a lex contractus or lex praepositionis, in that it expounded the terms of the contract of hire between shipper and merchants.75

The legal status of such a usage is uncertain, but somewhat illuminated by a very controversial text by the late second-century jurist Volusius Maecianus, the alleged author of a monograph on the lex Rhodia (D. 14.2.9). A petitioner writes to the emperor to complain about being robbed by islanders after a shipwreck. The emperor’s answer, based on an earlier ruling by Augustus, specifies that the lex Rhodia applies whenever it is not in conflict with Roman law. Taken at face value, the text indicates that legal vacuum could – or had to – be filled with existing usage, whatever they are or come from: foreign customs are better than nothing, and Roman lawmakers could not be expected to cover all situations. Maritime law was obviously permeable to external input.

Public law: regulation and exploitation

The combination of edictal law and jurisprudence, and the occasional adoption of international/local norms (laws and customs) certainly facilitated the development of commerce by providing the business community with adequate legal instruments and protection. However, there is another side to the coin, reflecting social and political concerns and fiscal necessities. This is where public law steps in.

Starting in the early Republican period, a series of leges fenebres tended to limit the rate of interest before banning interest altogether, though unsuccessfully (cf. above). Other public laws bearing on commercial activities regulated the food supply (leges

74 In the seventh cent., Isidorus of Seville (Etym. 5.17) speaks of Rhodian laws of maritime commerce as ‘antiquitus mercatorum usus.’

75 Zimmermann (1990), 406-12; Chevreau (2005); Aubert (2007a); Badoud (forthcoming).

 

21

 

annonariae/frumentariae), luxury consumption (leges sumptuariae), the occupation of agricultural land (leges agrariae), and taxation.76 The point was to preserve the social order and the political power of the elite. One such law, the plebiscitum Claudianum of 219/218 BC, reiterated in slightly different form by the lex Iulia repetundarum of 59 BC, barred senators and their sons from owning – though not operating – ships of large capacity (over 300 amphoras), excluding them from lucrative public contracts connected with the food supply and hampering the marketing of the produce of ever-growing agricultural estates.77 The ban may have been instrumental in developing the actio exercitoria and its extension to non¬dependent ship’s captains (cf. above). Alternatively, the activities of shippers (navicularii) drew the attention of imperial government officials and gave rise to an abundant legislation down to late Antiquity. At stake was the steadiness of the food supply of Rome and, since the fourth century, that of Constantinople. Organized in associations (collegia, corpora), they enjoyed at first privileges such as exemption from compulsory public services, and ended

up fulfilling a public service (onus publicum) even against their will.78

State control over economic activities was not limited to shipping and became a general phenomenon in the fourth and fifth centuries. Unsurprisingly, it affected trades connected with the food industry (bakers, meat sellers, etc.), but extended to other

commercial activities.79 It was mostly exercised through taxation: in the Republican and early Imperial periods, trade was subjected to all kinds of taxes (vectigalia), above all tolls and custom duties (portoria) at both municipal and imperial levels.80 It is difficult to estimate the impact of taxation on the volume of trade, but it is clear that the burden increased with time: from the reign of Constantine until 498, a special tax called collatio lustralis or chrysarguron was collected in gold and silver on behalf of the imperial treasury (sacrae largitiones) from merchants (negotiatores, mercatores) who therefore had to be registered

(negotiatorum matricula).81 By then, the time of laisser-faire and empiric promotion of commercial activities on the part of public authorities was long gone.

Traders were not only considered a fiscal golden hen by a needy government. The attitude of lawmakers towards them had changed. The preamble of Diocletian’s Price Edict (Nov./Dec. 301) accuses them in no uncertain terms of greed and selfishness, the cause of uncontrollable inflation, and threatens them and their agents (institores) with capital punishment unless they desist from speculating and abide by the law setting maximum

76 Rotondi (1912(1990), 92-100.

77 Liv. 21.63; Cic., 2Verr. 5.17.44-18.45; and Ps.-Paul, Sent. (Leiden frg. published in 1956) 3 (p. 5), ll. 7¬11; Aubert (2004), 166-68 and 178, n. 17; Tchernia (2007), with a telling calculation of the capacity of the ship.

78 Sirks (1991); de Salvo (1992).

79 Pistores, suarii, pecuarii, boarii, vinarii, etc. CI., Book 11; CTh., Books 10 and 14; Sirks (1991), 307-413; Földi (2001), 85-87.

80 France (2001), for Gaul; Cottier et al. (2008), for Asia Minor; and Delmaire (1989), 275-312, for late Antiquity.

81 CTh. 13.1; CI. 11.1 (abolition in the East); Jones (1964), 351 and 431-32; Delmaire (1989), 254-74, esp. 367, n. 41, with reference (for the matricula) to CTh. 16.2.15.1 (359 or 360); and P. Oxy. L 3577 (28 Jan. 342). Cf. also Minaud (2011), §298. Other taxes on trade are attested: siliquaticum, canon telonei and transmarinorum, cf. Jones (1964), ss.vv.

 

22

 

prices, possibly disconnected from market prices. A few years later, Lactantius claimed that the ill-advised imperial policy resulted in both slaughter and scarcity of goods.82 In spite of its failure and eventual repealing, the measure shows that soldiers, not traders, have the emperors’ ear.

Conclusion

To the question, was Roman business law designed for traders, we can propose a qualified answer: looking at the time of its development (Republic and Principate), there is no doubt that edictal law and jurisprudence jointly produced legal institutions of unprecedented efficiency, offering pragmatic solutions to practical problems, occasionally borrowed from subjected communities. This is true of private law. As for public law, it was concerned with social stability and fiscal necessity, not economic growth, and should be viewed as a permanent hindrance.

To the question, was Roman business law inspired by traders, the answer is less clear-cut. Roman magistrates and jurists invariably belonged to the elite. The dominant ideology would have liked us to think of it as a landed aristocracy, but the evidence suggests that senatorial, equestrian, and curial families were heavily –for senators, perhaps indirectly – involved in commercial activities, in spite of legal prohibition and social pressure. Because of the pragmatic nature of edictal law and the flexibility in its application warranted in juristic writings, it is difficult to exclude close and recurrent contacts between the business community and lawmakers, as Cicero’s indiscretion occasionally reveals. The sheer quantity and strong rhetorical taste of legal opinions preserved in the Digest cannot be allowed to blur the fact that Roman business law is mostly edictal law, explained and extended by the jurists.

To the question, was Roman business law enforced by traders or, at least, in special courts reserved for traders, the answer can only be a non liquet: there is not much evidence

for a Roman equivalent to the Greek emporikai dikai.83 Roman courts were composed of recuperatores, sometimes limited to a iudex unus or arbiter, and selected from a list of respected people, some of whom could have been negotiatores (Cic., 2Verr. 2.13.32-34). In the imperial period, jurisdiction passed to civil servants (praefecti, praesides, iudices) whose interests can hardly be confused with those of traders, and whose technical competence rested less on their personal legal expertise and practical experience than on the services of their staff. The evident worsening of the condition of traders in late Antiquity does not point toward the establishment of judicial privileges.

82 Edictum Diocletiani et collegarum de pretiis rerum venalium (ed. M. Giacchero, 1974), 1: 134-37, esp. ll. 64-136; Lact., De mortibus persecutorum 7.6-7. Cf. Corcoran (2000), 205-33.

83 Cohen (1973); and id. (2005), 300-2. The Roman evidence collected by Minaud (2011), §365 (Cic., 2Verr. 2.13.34; CIL XIV 2630; SHA, Alex. 33.2; CTh. 14.7.1 [397]; CI. 3.13.7 pr. [502]; Isid., Etym. 5.17) is inconclusive, with the possible exception of the passage from the Historia Augusta.

23

 

Roman Business Law: Bibliography

Andreau, J. (1987) La vie financière dans le monde romain. Les métiers de manieurs d’argent (IVe siècle av. J.-C. –IIIe siècle ap. J.-C.). Rome.

Andreau, J. (1999) Banking and Business in the Roman World. Cambridge.

Ankum, H. (2000) ‘Some aspects of maritime loans in old-Greek and in Roman law,’ in Velissaropoulou-Karakosta, I et al. (eds.), Mélanges I. Triantaphyllopoulos. Komotini, 293-306.

Aubert, J.-J. (1993) ‘Workshop managers,’ in Harris, W.V. (ed.), The Inscribed Economy. Production and Distribution in the Roman Empire in the Light of instrumentum domesticum. Ann Arbor, MI, 171-89.

Aubert, J.-J. (1994) Business Managers in Ancient Rome. A Social and Economic Study of Institores, 200 BC – AD 250. Leiden/New York/Cologne.

Aubert, J.-J. (1999a) ‘Les institores et le commerce maritime dans l’empire romain,’ Topoi 9: 145¬64.

Aubert, J.-J. (1999b) ‘La gestion des collegia: aspects juridiques, économiques et sociaux,’ CCG 10: 49-69.

Aubert, J.-J. (2003) ‘En guise d’introduction: contrats publics et cahiers des charges,’ in Aubert, J.-J. (ed.) Tâches publiques et entreprise privée. Neuchâtel/Genève, 1-25.

Aubert, J.-J. (2004) ‘The Republican economy and Roman law: regulation, promotion, or reflection?’ in Flower, H. I., The Cambridge Companion to the Roman Republic. Cambridge, 160-78.

Aubert, J.-J. (2005a) ‘Corpse disposal in the Roman colony of Puteoli: public concern and private enterprise,’ in Harris, W.V. – Lo Cascio, E. (eds.) Noctes Campanae. Studi di storia antica ed archeologia dell’Italia preromana e romana in memoria di Martin W. Frederiksen. Napoli, 141-57.

Aubert, J.-J. (2005b) ‘L’estampillage des briques et des tuiles: une explication juridique fondée sur une approche globale,’ in Bruun, C. (ed.), Interpretare i bolli laterizi di Roma e della Valle del Tevere: Produzione, storia economica e topografia. Rome, 53-59.

Aubert, J.-J. (2007a) ‘Dealing with the abyss: the nature and purpose of the Rhodian sea-law on jettison (Lex Rhodia de iactu, D 14.2) and the making of Justinian’s Digest,’ in Cairns, J. W. – du Plessis, P. (eds.), Beyond Dogmatics. Law and Society in the Roman World. Edinburgh, 157-72.

Aubert, J.-J. (2007b) ‘L’économie romaine et le droit de la représentation indirecte sous la République romaine,’ in Fides, Humanitas, Ius. Studii in onore di Luigi Labruna. Napoli, 215-30.

Aubert, J.-J. (2010) ‘Productive investment in agriculture: instrumentum fundi and peculium in the later Roman Republic,’ in Carlsen, J. – Lo Cascio, E. (eds.), Agricoltura e scambi nell’Italia tardo-repubblicana. Bari, 167-85.

Aubert, J.-J. (2011) ‘Vitia animi: tares mentales, psychologiques, caractérielles et intellectuelles des esclaves en droit romain,’ in Maffi, A. – Gagliardi, L. (eds.), I diritti degli altri in Grecia e a Roma. Sankt Augustin, 236-48. Aubert, J.-J. – Raepsaet, G. (2011), ‘Un mandat inscrit sur une sigillée argonnaise à Liberchies-Geminicacum,’ in L’Antiquité classique 80,139-56.Badoud, N. (forthcoming), La lex Rhodia de iactu, to be published in the Schweizerische Beiträge zur Altertumswissenschaft, based on his unpublished 2008 Neuchâtel doctoral dissertation, La cité de Rhodes. De la chronologie à l’histoire (Second part, 443-569).

Bianchini, M. (1989) ‘Diritto commerciale nel diritto romano,’ in Digesto delle Discipline privatistiche. Sezione commerciale IV. Torino, 320-33.

24

 

Bianchini, M. (2007) ‘Attività commerciali fra privato e pubblico in età imperiale,’ in Fides,

Humanitas, Ius. Studii in onore di Luigi Labruna. Naples, 423-38.

Biscardi, A. (1974) Actio pecuniae traiecticiae. Contributo alla dottrina delle clausole penali. Sec. ed.,

Torino.

Borkowski, A. – Du Plessis, P. (2005) Textbook on Roman law. Third ed., Oxford.

Brennan, T. C. (2000) The Praetorship in the Roman Republic. Oxford.

Capogrossi Colognesi, L. (1994) ‘Ius commercii, conubium, civitas sine suffragio. Le origini del

diritto internazionale privato e la romanizzazione delle comunità latino-campane,’ in

Corbino, A. (ed.), Le strade del potere. Catania, 3-64.

Carnazza, G. (1891) Il diritto commerciale dei Romani. Catania (non vidi).

Casson L. (1986) ‘New light on maritime loans: P. Vindob. G. 19792 (= SB VI 9571),’ in

Bagnall, R.S. – Harris, W.V. (eds.), Studies in Roman Law in Memory of A. Arthur Schiller.

Leiden/New York, 11-17.

Casson L. (1990) ‘New light on maritime loans: P. Vindob. G 40822,’ ZPE 84: 195-206.

Cerami, P. – Petrucci, A. (2010) Diritto commerciale romano. Profilo storico. Third ed., Torino.

Champaud, C. (1994) Le droit des affaires. Fifth ed., Paris.

Chevreau, E. (2005) ‘La lex Rhodia de iactu: un exemple de la réception d’une institution

étrangère dans le droit romain,’ TvR 73: 67-80.

Chiusi, T. J. (2001) Die actio de in rem verso im römischen Recht. Munich.

Chiusi, T. J. (2003) Contributo allo studio dell’editto ‘de tributoria actione’. Rome.

Churruca, J. de (2001) ‘Le commerce comme élément de civilisation dans la Géographie de

Strabon,’ RIDA 48: 41-56.

Cogliolo, P. (1922) Gli istituti commerciali nel diritto romano. Corso compilato dall’Avv. E. Fin!Zi e dallo

studente G.G. Traverso, Anno accademico 1921-1922. Genova.

Cohen, E. E. (1973) Ancient Athenian Maritime Courts. Princeton.

Cohen, E. E. (2005) ‘Commercial law,’ in Gagarin, M. – Cohen, D. (eds.), The Cambridge

Companion to Ancient Greek Law. Cambridge, 290-302.

Coppola Bisazza, G. (2003) Lo iussum domini e la sostitu!Zione nego!Ziale nell’esperien!Za romana.

Milano.

Corcoran, S. (2000), The Empire of the Tetrarchs: Imperial Pronouncements and Government, AD 284 

324. Rev. ed., Oxford.

Cottier, M. et al. (eds.) (2008) The Customs Law of Asia. Oxford.

Crook, J. A. (1967) Law and Life of Rome, 90 B.C. – A.D. 212. Ithaca, NY.

Daube, D. (1951) ‘The peregrine praetor,’ JRS 41: 66-70.

Delmaire, R. (1989) Largesses sacrées et res privata. L’aerarium impérial et son administration du IVe

au VIe siècle. Rome.

De Salvo, L. (1992) I corpora naviculariorum: economia privata e pubblici servi!Zi nell’impero romano.

Messina.

Di Porto, A. (1984) Impresa collettiva e schiavo ‘manager’ in Roma antica (II sec. a.C. – II sec. d.C.).

Milano.

Di Porto, A. (1997) ‘Il diritto commerciale romano. Una ‘zona d’ombra’ nella storiografia

romanistica e nelle riflessioni storico-comparative dei commercialisti,’ in Marrone, M. (ed.),

No!Zione, forma!Zione e interpreta!Zione del diritto dall’età romana alle esperien!Ze moderne. Ricerche dedicate

al Prof. Filippo Gallo. Napoli, 3: 413-52.

Fadda, C. (1903/1987) Istituti commerciali del diritto romani. Le!Zioni dettate nella R. Universitä di

Napoli nell’anno scolastico 1902-1903. Naples. (Repr. con una nota di lettura di L. Bove).

 

25

 

Ferenczy, E. (1969) ‘Die römisch-punischen Verträge und die Protohistorie des commercium,’ RIDA 16 (1969) 259-82.

Földi, A. (2001) ‘Eine alternative Annäherungsweise: Gedanken zum Problem des Handelsrechts in der römischen Welt,’ RIDA 48: 65-90.

France, J. (2001) Quadragesima Galliarum. L’organisation douanière des provinces alpestres, gauloises et germaniques de l’empire romain. Rome.

Frier, B. W. (1980) ‘Roman law and the wine trade: the problem of ‘vinegar’ sold as wine,’ ZRG 100: 257-95.

Gaudemet, J. (1982). Institutions de l’Antiquité. 2nd ed., Paris.

Gaurier, D. (2004) Le droit maritime romain. Rennes.

Gerkens, J.-F. (2003) ‘Chronique de la 55e session de la Société internationale Fernand de Visscher pour l’histoire des droits de l’Antiquité à Rotterdam (18 au 22 septembre 2001),’ RIDA 50: 489-516.

Giachi, C. (2001) ‘Storia dell’editto e struttura del processo in età pre-adrianea. Un’ipotesi di lavoro,’ in Atti del Convegno Processo civile e processo penale nell’esperienza giuridica del mondo antico in memoria di Arnaldo Biscardi, Siena, Certosa di Pontignano, 93-95 dicembre 2009. Rivista di diritto romano http://www.ledonline.it/rivistadirittoromano/attipontignano.html (12/01/09).

Giachi, C. (2005) Studi su Sesto Pedio. La tradizione, l’editto. Milano.

Goode, R. (2004) Commercial Law. Third ed. London.

Griffin, M. (1973) ‘The tribune C. Cornelius,’ JRS 63: 196-213.

Gröschler, P. (2002) Actiones in factum. Eine Untersuchung zur Klage-Neuschöpfung im nichtvertraglichen Bereich. Berlin.

Guarino, A. (1980) ‘La formazione dell’editto perpetuo,’ ANRW 2.13: 62-102.

Harris, W. V. (2003) ‘Roman governments and commerce, 300 B.C. – A.D. 300,’ in Zaccagnini, C. (ed.), Mercanti e politica nel mondo antico. Rome, 275-305.

Hoz Montoya, J. de la (2008), ‘Neron y el impuesto sobre la venta de esclavos,’ SDHI 74: 367¬395.

Humbert, M. (2005) ‘Il valore semantico e giuridico di vsvs nelle Dodici Tavole,’ in Humbert, M. (ed.), Le Dodici Tavole. Dai Decemviri agli Umanisti. Pavia, 377-400.

Huvelin, P. (1929) Etudes d’histoire du droit commercial romain (Histoire externe – Droit maritime). Paris.

Impallomeni, G. (1955) L’editto degli edili curuli. Padova.

Jakab, E. (2008) ‘Vertragspraxis und Bankgeschäfte im antiken Puteoli: TPSulp. 48 neu interpretiert,’ in Verboven, K. – Vandorpe, K. – Chankowski, V. (eds.), Pistoi dia tèn technèn. Bankers, Loans and Archives in the Ancient World. Studies in Honour of Raymond Bogaert. Leuven, 321-44.

Johnston, D. (1999) Roman Law in Context. Cambridge.

Jones, A. H. M. (1964) The Later Roman Empire, 284-602. A Social, Economic, and Administrative Survey. London/Norman, OK (repr. Baltimore 1986).

Jones, D. (2006) The Bankers of Puteoli. Finance, Trade and Industry in the Roman World. Stroud. Kaser, M. (1953) ‘Vom Begriff des commercium,’ in Studi in onore di V. Arangio-Ruiz. Naples, 2: 131-67.

Kaser, M. (1971-1975) Das römische Privatrecht. Second ed. Munich.

Kelly, D. – Holmes, A. – Hayward, R. (2005), Business Law. Fifth ed., Abingdon/New York. Kremer, D. (2005) ‘Trattato internazionale e legge delle Dodici Tavole,’ in Humbert, M. (ed.), Le Dodici Tavole. Dai Decemviri agli Umanisti. Pavia, 191-207.

 

26

 

Kudlien, F. (2001) ‘Mutator und permutatio als finanz- und handelstechnische Termini,’ MBAH

20.1: 48-54.

Labruna, L. (1994) ‘Il diritto mercantile dei Romani e l’espansionismo,’ in Corbino, A. (ed.), Le

strade del potere. Catania, 115-37.

Legeais, D. (2003) Droit commercial et des affaires. Fifteenth ed., Paris.

Litewski, W. (1973) ‘Römisches Seedarlehen,’ Iura 24: 112-83.

Lo Cascio, E. (2007) ‘The early Roman empire: the state and the economy,’ in Scheidel, W. –

Morris, I. – Saller, R. (eds.), The Cambridge Economic History of the Greco-Roman World.

Cambridge, 619-47.

Lucas, F.-X. (2005) Le droit des affaires. Paris.

Malanima, P. (2009) ‘Progresso o stabilità? il mercato nelle economie preindustriali,’ Studi storici

50: 5-23.

Maganzani, L. (2002) Pubblicani e debitori d’imposta. Ricerche sul titolo edittale de publicanis. Torino.

Maganzani, L. (2007) ‘L’editto provinciale alla luce delle Verrine: profili strutturali, criteri

applicativi,’ in Dubouloz, J. – Pittia, S. (eds.), La Sicile de Cicéron. Lectures des Verrines.

Besançon, 127-46.

Mantovani, D. (2000) ‘L’édit comme code,’ in Lévy, E. (ed.), La codification des lois dans

l’Antiquité. Actes du colloque de Strasbourg, 27-29 novembre 1997. Paris, 257-72.

Mayer-Maly, Th. (2003) ‘Commercium,’ TR 71: 1-6.

Meyer, E. A. (2004) Legitimacy and Law in the Roman World. Tabulae in Roman Belief and Practice.

Cambridge.

Meylan, Ph. (1959) ‘Permutatio rerum,’ in Ius et lex. Festgabe zum 70. Geburtstag von M. Gutzwiller.

Basel, 45-64 (repr. 1977).

Merryman, J. H. (1985) The Civil Law Tradition. An Introduction to the Legal Systems of Western

Europe and Latin America. Second ed. Stanford.

Miceli, M. (2001) Sulla struttura formulare delle actiones adiecticiae qualitatis. Torino.

Minaud, G. (2011), Les gens de commerce et le droit à Rome. Aix-en-Provence.

Nörr, D. (2005) ‘Osservazioni in tema di terminologia giuridica predecemvirale e di ius

mercatorum mediterraneo: il primo trattato cartaginese-romano,’ in Humbert, M. (ed.), Le

Dodici Tavole. Dai Decemviri agli Umanisti. Pavia, 147-89.

North, D. (1981) Structure and Change in Economic History. New York/London.

North, D. (1991) Institutions, Institutional Change and Economic Performance. Cambridge.

North, D. – Thomas, R.P. (1973) The Rise of the Western World: A New Economic History.

Cambridge.

Palazzolo, N. (1991) ‘Sulla lex Cornelia de edictis,’ Labeo 37.2: 242-45.

Peppe, L. (1991) ‘Note sull’editto di Cicerone in Cilicia,’ Labeo 37: 14-93.

Petrucci, A. (1991) Mensam exercere. Studi sull’impresa finanziaria romana (II secolo a.C. – metà del

III secolo d.C.). Napoli.

Petrucci, A. (2002) Profili giuridici delle attività e dell’organizzazione delle banche romane. Torino.

Pinna Parpaglia, P. (1987) Per una interpretazione della ‘lex Cornelia de edictis praetorum’ del 67 A.C.

Sassari.

Pinna Parpaglia, P. (1992) ‘Lex Cornelia de edictis, mutui feneratizi, certezza del diritto,’ Labeo

38.2: 372-78.

Pugliese, G. (1964) ‘Riflessioni sull’editto di Cicerone in Cilicia,’ in Synteleia V. Arangio-Ruiz.

Napoli, 2: 972-96 (= [1985] Scritti giuridici scelti. Napoli, 99-115).

 

27

 

Pugsley, D. (1974) ‘The Aedilician Edict,’ in Watson, A. (ed.), Daube Noster. Essays in Legal History for David Daube. Edinburgh/London, 253-64.

Reduzzi-Merola, F. (2001) ‘Ventes d’esclaves sur les marchés de Campanie,’ in Routes et marchés d’esclaves. 26e colloque du GIREA. Besançon, 321-25.

Rotondi, G. (1912) Leges publicae populi romani. Milano.

Rougé, J. (1966) Recherches sur l’organisation du commerce maritime en Méditerranée sous l’empire romain. Paris.

Santalucia, B. (1975) L’opera di Gaio ‘ad edictum praetoris urbani’. Milano.

Scardigli, B. (1991) I trattati Romano-Cartaginesi. Pisa.

Schleppinkhoff, A. M. M. (1996) Actio quod iussu. Die Geheissklage (und ihre Bedeutung für die Entwicklung des Stellvertretungsgedanken im 19. Jahrhundert. Diss. Köln.

Schuster, S. (2005) Das Seedarlehen in den Gerichtsreden des Demosthenes: mit einem Ausblick auf die weitere historische Entwicklung des Rechtsinstitutes: dáneion nautikón, fenus nauticum und Bodmerei. Berlin.

Serrao, F. (1954) La ‘iurisdictio’ del pretore peregrino. Milano.

Serrao, F. (1989) Impresa e responsabilità a Roma nell’età commerciale: forme giuridiche di un’economia-mondo. Ospedaletto.

Serrao, F. (2000) ‘Impresa, mercato, diritto. Riflessioni minime,’ in Lo Cascio, E. (ed.), Mercati permanenti e mercati periodici nel mondo romano. Atti degli Incontri capresi di storia dell’economia antica (Capri 13-15 ottobre 1997). Bari, 31-67.

Silver, M. (2007) ‘Roman economic growth and living standards: perceptions versus evidence,’ Ancient Society 37: 191-252.

Sirks, B. (1991) Food for Rome. Amsterdam.

Sirks, B. (2002) ‘Sailing in the off-season with reduced financial risk,’ in Aubert, J.-J. – Sirks, B. (eds.), Speculum iuris. Roman Law as a Reflection of Social and Economic Life in Antiquity. Ann Arbor, MI, 134-50.

Sirks, B. (forthcoming) ‘Law, commerce and finance in the Roman empire,’ unpublished paper presented at the Fourth OxREP conference, 1-4 October 2009.

Tchernia, A. (2007) ‘Le plebiscitum Claudianum,’ in Andreau, J. – Chankowski, V. (eds.), Vocabulaire et expression de l’économie dans le monde antique. Bordeaux, 253-78.

Terpstra, T. (2008) ‘Roman law, transaction costs and the Roman economy: evidence from the Sulpicii archive,’ in Verboven, K. – Vandorpe, K. – Chankowski, V. (eds.), Pistoi dia tèn technèn. Bankers, Loans and Archives in the Ancient World. Studies in Honour of Raymond Bogaert. Leuven, 345-69.

Wolff, H.-J. (1979) Das Problem der Konkurrenz von Rechtsordnungen in der Antike. Heidelberg. Zimmermann, R. (1990) The Law of Obligations. Roman Foundations of the Civilian Tradition. Oxford.

 

28

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

LAKIREDDY BALIREDDY COLLEGE OF ENGINEERING

(AUTONOMOUS)

(Approved by AICTE, Accredited by NBA,

Affiliated to JNTUK, Kakinada and ISO 9001: 2008 Certified)

ACADEMIC REGULATIONS,

COURSE STRUCTURE

AND

DETAILED SYLLABUS

M.TECH – SYSTEMS AND SIGNAL PROCESSING 

DEPARTMENT OF ELECTRONICS AND COMMUNICATION

ENGINEERING 

L.B.Reddy Nagar, MYLAVARAM – 521 230

Krishna District, Andhra Pradesh State

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 0

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

INDEX

Contents

1. Introduction

2. Programme Offered Post raduation Page No.

2

2

3. Eligibility Criteria for dmission 2

4. ard of M. ec Degree 3

5. Duration of t e Programme 3

6. Semester- ise distribution of credits 3

7. Distribution and eig tage of Mar s 3

8. ttendance Regulations Condonation 5

9. Minimum cademic Re uirements 6

10. Course Pattern 6

11. ard of rade 7

12. Minimum Instruction days 8

13. eneral 8

14. ransitory Regulations 8

15. Course Code and Course umbering Sc eme 9

16. Medium of Instruction 10

17. mendments to Regulations 10

18. rade Card 10

19. Conduct and Disci line 10

20. Mal ractices 12

21. ard of Ran 12

22. Course structure 14

23. Syllabus 16-42

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 1

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

1. INTRODUCTION

cademic Programmes of t e institute are governed by rules and regulations as

a roved by t e cademic Council, ic is t e ig est cademic body of t e

Institute.

ese academic rules and regulations are a licable to t e students admitted during academic year 2010-11 into first year of t o year Postgraduate rogramme offered by t e college leading to Master of ec nology M. ec degree.

1.1 a ireddy alireddy College of Engineering, Mylavaram, an autonomous

institution, follo s Semester attern for all t o years of its Postgraduate M. ec rogramme it internal and e ternal evaluation.

1.2 Semester Pattern : Eac academic year s all be divided into t o semesters, eac of 20 ee s duration, including instruction, evaluation, etc. Eac semester consists of a minimum of 90 instruction days it at least 35 to 40 contact eriods er ee .

2. PROGRAMME OFFERED (POST GRADUATE)

Master of ec nology M. ec .

3. ELIGIBILITY CRITERIA FOR ADMISSION

e eligibility criteria for admission into irst year M. ec rogramme s all be as

mentioned belo :

i. dmissions to t e above rogram s all be made sub ect to t e eligibility, ualifications and s eciali ation rescribed by t e IC E from time to time.

ii. dmissions s all be made on t e basis of merit ran obtained by t e

ualifying candidate at E e amination or an entrance test conducted by

t e university sub ect to reservations rescribed by t e niversity/State government from time to time.

4. AWARD OF M.Tech DEGREE

student ill be declared eligible for t e a ard of t e M. ec Degree if e/s e fulfills t e follo ing academic regulations:

i. Pursued a course of study for not less t an t o academic years and not more t an four academic years.

ii. e student as to fulfill all t e academic re uirements i.e. Registered for 88 credits and as to secure all t e 88 credits it minimum grade oints.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 2

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

5. DURATION OF THE PROGRAMME

Students, o fail to fulfill all t e academic re uirements for t e a ard of t e degree it in four academic years from t e year of t eir admission, s all forfeit t eir seat in M. ec course.

6. SEMESTER –WISE DISTRIBUTION OF CREDITS:

Year First Semester

Credits Second Semester

Credits Total Credits

irst ear 3028

Mini Pro ect 2 Credits 3028

erm Pa er 2 Credits 60

Second ear Dissertation or 20

ec nical seminar-8 28

TOTAL 88

Table .1 Semester wise Credits Distribution 

7. DISTRIBUTION AND WEIGHTAGE OF MARKS:

i In I-Semester and II- Semesters, t e course of study consists of 6 t eory sub ects 2

laboratories or 6 t eory sub ects 1 aboratory. Ho ever, t e final year ill be on dissertation or and ec nical seminar only.

ii e erformance of a student in eac semester s all be evaluated sub ect ise it a ma imum of 100 mar s for t eory and 100 mar s for aboratory Courses. In addition, Mini ro ect, ec nical seminar, and dissertation or s all be evaluated for 50, 50 and 200 mar s res ectively.

iii or t eory sub ects t e distribution s all be 40 mar s 35 for est and 5 for

attendance for Internal E aminations and 60 mar s for t e End Semester E amination.

iv or t eory sub ects, during t e semester t ere s all be t o internal mid term

e aminations, for duration of 90 minutes. irst internal mid term e amination to be conducted in first and second units and t e second internal mid term

e amination to be conducted in t ird, fourt and fift units of eac sub ect.

Ho ever,75 eig tage for t e best and 25 for t e ot er internal mid term

e amination s all be considered for a arding internal mar s.

v e internal mid term e amination uestion a er s ould be for 35 mar s. Out of five

uestions given, student as to ans er any t ree uestions.

vi or aboratory courses, t ere s all be a continuous evaluation during t e semester

for 40 sessional mar s and 60 end semester e amination mar s. Of t e 40 mar s for internal, 15 mar s s all be a arded for day-to-day or and 15 mar s to be

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 3

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

a arded by conducting Internal laboratory test and 5 mar s for record or and 5 mar s for attendance. e end semester e amination s all be conducted by an e ternal e aminer and t e concerned internal e aminer.

vii Mini ro ect s all be submitted in re ort form and s ould be resented before t e

committee, ic s all be evaluated for 50 mar s. e committee consists of t e

Head of t e de artment, t e Su ervisor of mini ro ect and a senior faculty member oft e de artment. ere s all be no internal mar s for mini ro ect. e student as to secure minimum 50 mar s to be declared successful.

viii ere s all be ec nical seminar in second year. or t e seminar, t e student s all

collect t e information on a s eciali ed to ic and re are a tec nical re ort,

s o ing is understanding over t e to ic, and submit to t e de artment, ic

s all be evaluated by t e De artment committee consisting of Head of t e de artment, Seminar su ervisor and a senior faculty member. e seminars all be evaluated for 50 mar s based on is/ er resentation and t e submitted

re ort. ere s all be no e ternal e amination for seminar. e student as to

secure minimum 50 mar s to be declared successful.

i Dissertation Work: 

a. De artmental Dissertation Revie Committee DDRC s all be constituted it

t e Head of t e De artment as t e c airman and t o senior faculty as members to su ervise t e roceedings of t e dissertation or from t e time of allotment to submission.

b. Registration of Pro ect or :

student is ermitted to register for t e ro ect or after satisfying t e

attendance re uirement of all t e courses t eory and ractical courses u to

Second Semester. candidate as to submit, in consultation it is/ er

dissertation su ervisor, t e title, ob ectives and lan of action of is/ er dissertation or to t e DDRC for its a roval. Only after obtaining t e a roval of DDRC, t e student can initiate t e dissertation or .

c. e duration of t e dissertation is for t o semesters Second ear .

d. our co ies of t e dissertation re ort, certified by t e su ervisor s all be submitted tot e Institute.

e. Out of a total 200 mar s for t e dissertation or , 50 mar s s all be for Internal Evaluation and 150 mar s for t e End Semester E amination. e End Semester E amination viva-voce or defence s all be conducted by t e committee consisting of an E ternal E aminer, Head of t e De artment and t e dissertation su ervisor. E ternal e aminer ill be selected by t e Princi al/Director out of

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 4

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

t ree member anel submitted by t e Head of t e De artment, o are eminent

in t at field of study. or ma imum of 18 students, one dissertation evaluation committee as to be formed.

f. If t e or is not satisfactory, and if t e student is failed to secure t e minimum

50 of total mar s , t e student s all revise and resubmit t e dissertation re ort

before t ree mont s. If e/s e fails to get a satisfactory re ort again, t e dissertation s all be summarily re ected.

g. e to ics for mini ro ect and dissertation or s all be different from eac

ot er. e evaluation of dissertation or s all be conducted at t e end of t e

Second year.

. e student as to clear all t e sub ects of M. ec course to attend t e Viva

Voce or Defence of is / er Dissertation or .

i. e dissertation internal evaluation s all be on t e basis of continuous revie by

t e DDRC on t e rogress of t e dissertation or .

8. ATTENDANCE REGULATIONS & CONDONATION:

i. student s all be eligible to a ear for end semester e aminations, if ac uired a minimum of 75 of attendance in aggregate of all t e sub ects.

ii. Condonation of s ortage of attendance in aggregate u to 10 on medical grounds 65 and above and belo 75 in eac semester may be granted by t e College cademic Committee C C . Ho ever, t e sub ect of granting is urely at t e discretion of t e College cademic Committee or Com etent

ut ority.

iii. Student ill not be romoted to t e ne t semester unless e/s e satisfies t e

attendance re uirement of t e resent semester as a licable. ey may see re-admission for t at semester as and en offered ne t.

iv. Due eig tage in eac of t e sub ects s all be given to t e attendance. Mar s

not e ceeding 5 s all be given to all suc candidates o satisfy t e follo ing

criteria

Table .2 Marks weightage for Attendance

Percentage of attendance Mar s

90 5

85 to 90 4

80 to 85 3

75 to 80 2

75 1

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 5

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

v. S ortage of ttendance belo 65 in aggregates all in no case be condoned.

vi. Students ose s ortage of attendance is not condoned in any semester are not eligible to ta e t eir end Semester e aminations of t at articular semester and t eir registration for e amination s all stand cancelled.

vii. sti ulated fee s all be ayable to ards condonation of s ortage of attendance.

viii. ttendance may also be condoned for t ose o artici ate in restigious

s orts, co- and e tra-curricular activities rovided t eir attendance is in t e minimum rescribed range for t e ur ose and recommended by t e concerned aut ority.

9. MINIMUM ACADEMIC REQUIREMENTS:

e follo ing academic re uirements ave to be satisfied in addition to t e attendance re uirements mentioned in Item o.8.

i students all be deemed to ave secured t e minimum academic re uirement in

a sub ect if e/s e secures a minimum of 40 of mar s e clusively in t e end semester e amination and a minimum aggregate of 50 of t e total mar s in t e end semester e amination and internal evaluation ta en toget er. 

ii student ill be romoted to second year, if e/s e secures t e minimum

attendance re uirement.

Students o fail to earn 88 credits as indicated in t e course it in

four academic years from t e year of t eir admission s all forfeit t eir seat in M. ec course and t eir admission s all stand cancelled.

10. COURSE PATTERN:

e entire course of study is of t o academic years. Eac academic year s all ave

t o semesters.

Student eligible to a ear for t e end e amination in a sub ect, but absent at it

or as failed in t e end e amination may a ear for t at sub ect only as and en it

is conducted.

ll admitted students are to study 4 electives during t eir course of t o year study at t e institute. e follo ing s all be t e rogramme of study of electives.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 6

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

ear Semester o. of electives

1 1 2

1 2 2

inal year o semesters is com letely reserved for Dissertation or and

ec nical seminar.

en a student is detained due to s ortage of attendance/any ot er reason, e

may be re-admitted en t e semester is offered after fulfillment of academic

regulations. ereas, t e academic regulations old good it t e regulations

e/s e first admitted.

11. AWARD OF GRADE:

fter a student as satisfied t e re uirement rescribed for t e com letion of t e rogramme and is eligible for t e a ard of M. ECH Degree e/s e s all be laced in one of t e follo ing four grades. e a ard of t e degree is on a grade oint of scale

4. e grade oints are a arded as follo s:

CGPA Division of Pass

3.0 irst Class it Distinction

2.0 and 3.0 irst division

1.6 and 2.0 Pass division

1.6 ail

ased on t e erformance of t e candidate, e follo ing s all be t e criteria for t e

a ard of letter grades at t e end of eac semester in t e sub ects in ic t e

candidate a eared for t e e amination

Mar s

Scored rades rade

Points

90 S 4.00

85 to 90 3.67

80 and 85 3.33

75 and 80 3.00

70 and 75 2.67

65 and 70 C2.33

60 and 65 C 2.00

55 and 60 D 1.67

50 and 55 E 1.33

50 0

Table .3. Marks,Grades and Grade Points

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 7

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

11.1 Calculation of Semester Grade Points Average (SGPA)

e erformance of eac student at t e end oft e eac semester is indicated in terms of P . e S P is calculated as belo :

 

ere CR Credits of a course

GP rade oints a arded for a course

SGPA (Semester Grade Point Average) is calculated for t e candidates

o assed all t e courses in t at semester.

11.2 Calculation of Cumulative Grade Point Average (CGPA) for Entire Programme.

e C P is calculated as belo :

 

for entire rogramme

ere CR Credits of a course

GP rade oints a arded for a course

12. MINIMUM INSTRUCTION DAYS:

e minimum instruction for eac semesters all be 90 instruction days e cluding e amination days.

13. GENERAL:

a. ere t e ords e im is , occur in t e regulations, t ey include

s e , er , ers .

b. e academic regulations s ould be read as a ole for t e ur ose of any

inter retation.

c. In t e case of any doubt or ambiguity in t e inter retation of t e above rules, t e decision of t e Director is final.

d. e Institute may c ange or amend t e academic regulations or syllabi at any time and t e c anges or amendments made s all be a licable to all t e students it effect from t e dates notified.

14. TRANSITORY REGULATIONS

14.1 candidate, o is detained or discontinued in t e year/semester, on

readmission s all be re uired to do all t e courses in t e curriculum rescribed for

suc batc of students in ic t e student oins subse uently. Ho ever, e em tion

ill be given to t ose candidates o ave already assed in suc courses, ic

e/s e ad assed in t e earlier semester s e/s e as originally admitted into.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 8

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

14.2 student o is follo ing t e , a inada curriculum, detained due to

lac of academics/attendance at t e end of a semester of an academic year, s all oin it t e autonomous batc at t e a ro riate semester. Suc candidates s all be re uired to ass in all t e courses in t e rogramme rescribed by concerned OS for suc batc of students, to be eligible for t e a ard of degree. Ho ever,

e em tion ill be given to all t ose courses of t e semesters of t e batc , ic

t e candidate oins no , ic e/s e ad assed earlier. e student as to clear

all is bac log sub ects by a earing t e su lementary e aminations, conducted by

, a inada and utonomous stream for t e a ard of Degree. e mar s

secured by t e students in - a inada attern ill be converted in to

a ro riate grade oints as er t e autonomous grading system and t e class ill be

a arded based on t e academic erformance of a student in t e entire 2 years as er t e guidelines of autonomous Pattern.

14.3 e concerned oard of Studies s all give t e guidelines regarding Course E uivalence and Course E em tions from time to time.

15. COURSE CODE AND COURSE NUMBERING SCHEME:

Course umbers are denoted by 5 digit uni ue al a numeric c aracters. irst t o digits are M , t at describes t e Course name i.e. Master of ec nology.

Third digit re resents semester of offering as mentioned in able o. 4.

THIRD DIGIT DESCRIPTION

1 irst Semester

2 Second Semester

3 ird Semester

4 ourt Semester

Table 4: Third digit description 

Fourth digit re resents course ty e, as er able o. 5

O R H DI I DESCRIP IO

0 eory course

5 ab course/ot er t an t eory sub ect

Table 5 : Course type description 

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 9

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 10

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

c e follo ing acts of omission and/or commission s all constitute gross violation of

t e code of conduct and are liable to invo e disci linary measures it regard to ragging.

i. ac of courtesy and decorum indecent be avior any ere it in or outside t e cam us.

ii. illful damage or distribution of alco olic drin s or any ind of narcotics tot e fello students/citi ens.

d Possession, consum tion or distribution of alco olic drin s or any ind of narcotics or allucinogenic drugs.

e Mutilation or unaut ori ed ossession of library boo s.

f oisy and unseemly be avior, disturbing studies of fello students.

g Hac ing in com uter systems suc as entering into ot er erson s areas it out

rior ermission, mani ulation and/or damage of com uter ard are and soft are or any ot er cyber crime etc.

sage of camera cell ones in t e cam us.

i Plagiarism of any nature.

ny ot er act of gross indisci line as decided by t e academic council from time to time.

Commensurate it t e gravity of offense, t e unis ment may be re rimand, fine, e ulsion from t e institute / ostel, debarment from an e amination, disallo ing t e use of certain facilities of t e Institute, rustication for a s ecified eriod or even outrig t e ulsion from t e Institute, or even anding over t e case to a ro riate

la enforcement aut orities or t e udiciary, as re uired by t e circumstances.

l or an offence committed in i a ostel ii a de artment or in a class room and iii else ere, t e c ief arden, t e Head of t e De artment and t e rinci al res ectively, s all ave t e aut ority to re rimand or im ose fine.

m Cases of ado tion of unfair means and/or any mal ractice in an e amination s all be re orted to t e rinci al for ta ing a ro riate action.

n ll cases of serious offence, ossibly re uiring unis ment ot er t an re rimand,

s all be re orted to t e cademic council.

o e Institute evel Standing Disci linary ction Committee constituted by t e

academic council, s all be t e aut ority to investigate t e details of t e offence, and recommend disci linary action based on t e nature and e tent of t e offence committed.

e Princi al s all deal it any academic roblem, ic is not covered under

t ese rules and regulations, in consultation it t e Programmes Committee in an a ro riate manner, and subse uently suc actions s all be laced before t e

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 11

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

academic council for ratification. ny emergency modification of regulation, a roved by t e academic council earlier, s all be re orted to t e academic council for ratification.

“Grievance and Redressal Committee” (General) constituted by t e rinci al

s all deal it all grievances ertaining to t e academic/administrative /disci linary matters.

r ll t e students must abide by t e code and conduct rules of t e college.

21. MALPRACTICES

a e Princi al s all refer t e cases of mal ractices in internal assessment tests and

Semester-End E aminations, to a Mal ractice En uiry Committee, constituted by im/ er for t e ur ose. Suc committee s all follo t e a roved scales of

unis ment. e Princi al s all ta e necessary action, against t e erring students

based on t e recommendations of t e committee.

b ny action on t e art of candidate at an e amination trying to get undue

advantage in t e erformance at e aminations or trying to el anot er, or derive

t e same t roug unfair means is unis able according to t e rovisions

contained ereunder. e involvement of t e Staff, o are in c arge of

conducting e aminations, valuing e amination a ers and re aring/ ee ing records of documents relating to t e e aminations in suc acts inclusive of

roviding incorrect or misleading information t at infringe u on t e course of natural ustice to one and all concerned at t e e amination s all be vie ed seriously and recommended for a ard of a ro riate unis ment after t oroug en uiry.

22. AWARD OF RANK

e ran s all be a arded based on t e follo ing:

22.1 Only suc candidates o ass t e inal year e amination at t e

end of t e t ird academic year after admission as regular final year

students along it t e ot ers in t eir batc and become eligible for

t e a ard of t e Degree s all be eligible for t e a ard of ran .

Candidates, o loose one or more years of study for any reason

atsoever are not eligible for t e a ard of ran .

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 12

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

22.2 Ran s s all be a arded in eac branc of study for t e to five

students a earing for t e Regular e ternal E aminations.

22.3 ard of ri es, sc olars i s, or any ot er Honors s all be based on

t e ran secured by a candidate, consistent it t e desire of t e

Donor, erever a licable.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 13

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

M.TECH(ECE – SYSTEMS AND SIGNAL PROCESSING) - COURSE STRUCTURE

I-SEMESTER

Code No. Name of the Course Scheme of Instruction Scheme of Examination Total credits

Periods per Week Maximum Marks

Lecture Tutorial Lab. Internal External

MEC101 dvanced Digital Signal Processing 4 -- 40 60 100 4

MEC102 ransform ec ni ues 4 -- 40 60 100 4

MEC103 V SI ec nology and Design 4 -- 40 60 100 4

MEC104 Microcontrollers or Embedded System Design 4 -- 40 60 100 4

MEC1051

MEC1052 ELECTIVE – I 4 - - 40 60 100 4

DSP Processers rc itecture

Image and Video Processing

MEC1061

MEC1062 ELECTIVE – II 4 - - 40 60 100 4

Radar Signal Processing

io Medical Signal Processing

MEC151 Seminar -- 3 50 -- 50 2

MEC152 dvanced Digital Signal Processing ab -- 3 40 60 100 2

TOTAL 24 - 6 330 420 750 28

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 14

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

II-SEMESTER

Code No. Name of the Course Scheme of Instruction Scheme of

Examination Total credits

Periods per Week Maximum Marks

Lecture Tutorial Lab Internal External

MEC201 da tive Signal Processing 4 -- -- 40 60 100 4

MEC202 S eec Processing 4 -- -- 40 60 100 4

MEC203 SOC rc itecture 4 -- -- 40 60 100 4

MEC204 Coding eory and ec ni ues 4 -- -- 40 60 100 4

MEC2051

MEC2052 ELECTIVE – III 4 -- -- 40 60 100 4

CP D P rc itectures nd lications

Design for estability

MEC2061

MEC2062 ELECTIVE – IV 4 -- -- 40 60 100 4

ireless Communication and et or s

V SI Signal Processing

MEC207 Seminar -- -- 3 50 -- 50 2

MEC208 dvanced EC D ab -- -- 3 40 60 100 2

TOTAL 24 -- 6 330 420 750 28

III & IV SEMESTERS

Code No. Name of the Course Scheme of Instruction Scheme of Examination Total credits

Periods per Week Maximum Marks

Lecture Tutorial Lab. Internal External

MEC351 ec nical Seminar -- -- 6 50 50 8

MEC352 Dissertation -- -- 15 50 150 200 24

TOTAL -- -- 21 100 150 250 32

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 15

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

MEC101: ADVANCED DIGITAL SIGNAL PROCESSING

Lecture : 4 Periods/week Internal Marks : 40

External Marks : 60

Credits : 4 External Examination : 3 Hrs

UNIT - I

Review of DFT, FFT, IIR Filters, FIR Filters, Multirate Signal Processing: Introduction, Decimation by a factor D, Inter olation by a factor I, Sam ling rate conversion by a rational factor I/D, Multistage Im lementation of Sam ling Rate Conversion, ilter design

Im lementation for sam ling rate conversion, lications of Multirate Signal Processing

UNIT - II 

Non-Parametric methods of Power Spectral Estimation: Estimation of s ectra from finite

duration observation of signals, on- arametric Met ods: artlett, elc lac man

u ey met ods, Com arison of all on-Parametric met ods

UNIT - III 

Parametric Methods of Power Spectrum Estimation: utocorrelation Its Pro erties,

Relation bet een auto correlation model arameters, R Models - ule- a er urg

Met ods, M RM models for o er s ectrum estimation.

UNIT - IV

Linear Prediction : or ard and ac ard inear Prediction or ard inear Prediction,

ac ard inear Prediction, O timum reflection coefficients for t e attice or ard and ac ard Predictors. Solution of t e ormal E uations: evinson Durbin lgorit m, Sc ur lgorit m. Pro erties of inear Prediction ilters

UNIT - V

Finite Word Length Effects: nalysis of finite ord lengt effects in i ed- oint DSP

systems i ed, loating Point rit metic DC uanti ation noise signal uality inite

ord lengt effect in IIR digital ilters inite ord-lengt effects in algorit ms.

TEXTBOOKS

1. Digital Signal Processing: Princi les, lgorit ms lications - . .Proa is

t

D. .Manolo is, 4 ed., PHI.

2. Discrete ime signal rocessing - lan V O en eim Ronald Sc affer, PHI.

3. DSP Pratical roac Emmanuel C.Ifeac er, arrie. .ervis,

2 ed., Pearson Education.

REFERENCES:

1. Modern s ectral Estimation : eory lication S. M . ay, 1988, PHI.

2. Multirate Systems and ilter an s P.P.Vaidyanat an Pearson Education

3. Digital Signal Processing S.Saliva anan, .Vallavara , C. nana riya, 2000, MH

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 16

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

MEC102: TRANSFORM TECHNIQUES

Lecture : 4 Periods/week Internal Marks : 40

External Marks : 60

Credits : 4 External Examination : 3 Hrs

-------------------------------------------------------------------------------------------------------------- 

UNIT – I

Review of Transforms: Signal s aces, conce t of convergence, Hilbert s aces for energy

signals, ourier basis, -failure of -need for time-fre uency analysis, s ectrogram lot 

ase s ace lot in time-fre uency lane, Continuous ,D , Discrete ourier Series

and ransforms, - ransform, relation bet een C -D , D -D S,D S-D ,

DC 1D 2D, als ,Hadamard, Haar, Slant, ,Hilbert ransforms definition,

ro erties and a lications

UNIT – II

CWT & MRA: ime-fre uency limitations, tiling of time-fre uency lane for S ,

Heisenberg uncertainty rinci le, S ort time ourier ransform S analysis, s ort

comings of S , eed for avelets- avelet asis- Conce t of Scale and its relation it fre uebcy , Continuous time avelet ransform E uation- Series E ansion using avelets 

C - eed for scaling unction- Multi resolution analysis, iling of time scale lane for

C .Im ortant avelets : Haar, Me ican Hat Meyer, S annon, Daubec ies.

UNIT – III 

Multirate Systems, Filter Banks and DWT.

asics of Decimation and Inter olation in time fre uency domains, o-c annel ilter

ban , Perfect Reconstruction Condition, Relation s i bet een ilter an s and avelet

basis, D ilter an s or Daubec ies avelet unction

UNIT – IV

Special Topics: avelet Pac et ransform Multidimensional avelets, i-ort ogonal basis¬-s lines, ifting Sc eme of avelet eneration, Multi avelets

UNIT – V

Applications of Transforms

Signal Denoising, Subband Coding of S eec and Music, Signal Com ression - se of

DC , D , , 2-D D , ractal Signal nalysis.

TEXT BOOKS

1. undamentals of avelets- eory, lgorit ms and lications , aideva C

os ami, ndre C an, o n iley Sons, Inc, Singa ore, 1999.

2. avelet ransforms-Introduction t eory and a lications-Rag uveer M.Rao and it S. o ardi ar, Pearson edu, sia, e Del i, 2003.

3. Insig t into avelets from eory to ractice , Soman. .P, Ramac andran. .I, Printice Hall India, irst Edition, 2004.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 17

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

REFERENCES

1. avelets and sub-band coding , Vetterli M. ovacevic, P I, 1995.

2. Introduction to avelets and avelet ransforms , C. Sydney urrus, PHI, irst Edition, 1997.

3. avelet our of Signal Processing , Ste en . Mallat,. cademic Press, Second Edition,

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 18

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

MEC103: VLSI TECHNOLOGY AND DESIGN

Lecture : 4 Periods/week Internal Marks : 40

External Marks : 60

Credits : 4 External Examination : 3 Hrs

-------------------------------------------------------------------------------------------------------------- 

UNIT - I

Review of Microelectronics and Introduction to MOS Technologies: MOS, CMOS, iCMOS ec nology, rends nd Pro ections. asic Electrical Pro erties of MOS, CMOS BiCMOS Circuits: Ids-Vds relationships, Threshold Voltage Vt, Gm, Gds and ωo, Pass

ransistor, MOS, CMOS i CMOS Inverters, u/ d, MOS ransistor circuit model,

atc -u in CMOS circuits.

UNIT - II 

LAYOUT DESIGN AND TOOLS: ransistor structures, ires and Vias, Scalable Design rules, ayout Design tools.

LOGIC GATES & LAYOUTS: Static Com lementary ates, S itc ogic, lternative ate

circuits, o o er gates, Resistive and Inductive interconnect delays.

UNIT - III 

COMBINATIONAL LOGIC NETWORKS: ayouts, Simulation, et or delay, Interconnect design, Po er o timi ation, S itc logic net or s, ate and et or testing.

UNIT - IV

SEQUENTIAL SYSTEMS: Memory cells and rrays, Cloc ing disci lines, Design, Po er o timi ation, Design validation and testing.

UNIT – V

FLOOR PLANNING & ARCHITECTURE DESIGN: loor lanning met ods, off-c i

connections, Hig -level synt esis, rc itecture for lo o er, SOCs and Embedded CP s,

rc itecture testing.

TEXT BOOKS

1. Essentials of V SI Circuits and Systems, . Es rag ian Es rag ian. D, .Puc nell, 2005, PHI.

2. Modern V SI Design - ayne olf, 3rd ed., 1997, Pearson Education.

REFERENCES

nd

1. Princi als of CMOS V SI Design .H.E este, .Es rag ian, 2 ed., disson

esley.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 19

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

MEC104: MICROCONTROLLERS FOR EMBEDDED SYSTEM DESIGN

Lecture : 4 Periods/week Internal Marks : 40

External Marks : 60

Credits : 4 External Examination : 3 Hrs

-------------------------------------------------------------------------------------------------------------- 

UNIT – I

Introduction to Embedded Systems Overvie of Embedded Systems, Processor Embedded into a system, Embedded Hard are nits and Devices in system, Embedded Soft are, Com le System Design, Design Process in Embedded System, ormali ation of System Design, Classification of Embedded Systems.

UNIT – II

Microcontrollers and Processor rc itecture Interfacing 8051 rc itecture, In ut/Out ut

Ports and Circuits, E ternal Memory, Counters and imers, PIC Controllers. Interfacing Processor 8051, PIC , Memory Interfacing, I/O Devices, Memory Controller and Memory arbitration Sc emes.

UNIT - III

Embedded RISC Processors Embedded System-on C i Processor PSOC

Programmable System-on-C i arc itectures, Continuous imer bloc s, S itc ed

Ca acitor bloc s, I/O bloc s, Digital bloc s, Programming of PSOC, Embedded RISC

Processor arc itecture RM Processor arc itecture, Register Set, Modes of o eration and

overvie of Instructions

UNIT - IV

Interru ts Device Drivers E ce tions and Interru t andling Sc emes Conte t Periods

for Conte t S itc ing, Deadline interru t latency. Device driver using Interru t Service

Routine, Serial ort Device Driver, Device drivers for Internal Programmable timing devices

UNIT – V

et or Protocols Serial communication rotocols, Et ernet Protocol, SDM , C annel IDM , E ternal us Interface

TEXT BOOKS

nd

1. Embedded Systems - rc itecture Programming and Design Ra amal, 2 ed., 2008, MH.

2. PIC Microcontroller and Embedded Systems Mu ammad li Ma idi, Rolin

D.Mc inaly, Danny Causy PE.

3. Designers uide to t e Cy ress PSOC Robert s y, 2005, Elsevier.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 20

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

REFERENCES

1. Embedded Microcom uter Systems, Real ime Interfacing onat an . Valvano roo es / Cole, 1999, omas earning.

2. RM Systems Develo ers uides-Design O timi ing System Soft are - ndre . Sloss, Dominic Symes, C ris rig t, 2004, Elsevier.

3. Designing it PIC Microcontrollers- o n . Peatman, 1998, PH Inc.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 21

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

MEC1051: DIGITAL SIGNAL PROCESSORS AND ARCHITECTURE

Lecture : 4 Periods/week Internal Marks : 40

External Marks : 60

Credits : 4 External Examination : 3 Hrs

-------------------------------------------------------------------------------------------------------------- 

UNIT – I

INTRODUCTION TO DIGITAL SIGNAL PROCESING

Introduction, Digital signal- rocessing system, e sam ling rocess, Discrete time

se uences. Discrete ourier ransform D and ast ourier ransform , inear

time-invariant systems, Digital filters, Decimation and inter olation, nalysis and Design tool

for DSP Systems M , DSP using M .

COMPUTATIONAL ACCURACY IN DSP IMPLEMENTATIONS

umber formats for signals and coefficients in DSP systems, Dynamic Range and Precision,

Sources of error in DSP im lementations, /D Conversion errors, DSP Com utational errors, D/ Conversion Errors, Com ensating filter.

UNIT - II 

ARCHITECTURES FOR PROGRAMMABLE DSP DEVICES

asic rc itectural features, DSP Com utational uilding loc s, us rc itecture and

Memory, Data ddressing Ca abilities, ddress eneration nit, Programmability and Program E ecution, S eed Issues, eatures for E ternal interfacing.

UNIT - III 

EXECUTION CONTROL AND PIPELINING

Hard are loo ing, Interru ts, Stac s, Relative ranc su ort, Pi elining and Performance, Pi eline De t , Interloc ing, ranc ing effects, Interru t effects, Pi eline Programming models.

PROGRAMMABLE DIGITAL SIGNAL PROCESSORS

Commercial Digital signal- rocessing Devices, Data ddressing modes of MS320C54

DSPs, Data ddressing modes of MS320C54 Processors, Memory s ace of

MS320C54 Processors, Program Control, MS320C54 instructions and

Programming, On-C i Peri erals, Interru ts of MS320C54 rocessors, Pi eline

O eration of MS320C54 Processors.

UNIT - IV

IMPLEMENTATIONS OF BASIC DSP ALGORITHMS

e Q-notation, IR ilters, IIR ilters, Inter olation ilters, Decimation ilters, PID

Controller, da tive ilters, 2-D Signal Processing.

IMPLEMENTATION OF FFT ALGORITHMS

n lgorit m for D Com utation, utterfly Com utation, Overflo and scaling, it 

Reversed inde generation, n 8-Point im lementation on t e MS320C54 ,

Com utation of t e signal s ectrum.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 22

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

UNIT - V

INTERFACING MEMORY AND I/O PERIPHERALS TO PROGRAMMABLE DSP DEVICES Memory s ace organi ation, E ternal bus interfacing signals, Memory interface, Parallel I/O interface, Programmed I/O, Interru ts and I/O, Direct memory access DM . Multic annel buffered serial ort Mc SP , Mc SP Programming, a CODEC interface circuit, CODEC rogramming, CODEC-DSP interface e am le.

TEXT BOOKS

1. Digital Signal Processing vtar Sing and S. Srinivasan, omson Publications, 2004.

2. DSP Processor undamentals, rc itectures eatures a sley et al. 2000, S. C and Co.

REFERENCES

1. Digital Signal Processors, rc itecture, Programming and lications . Ven ataramani and M. as ar, 2002, MH.

2. Digital Signal Processing onat am Stein, 2005, o n iley.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 23

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

MEC1052: IMAGE AND VIDEO PROCESSING

Lecture : 4 Periods/week Internal Marks : 40

External Marks : 60

Credits : 4 External Examination : 3 Hrs

-------------------------------------------------------------------------------------------------------------- 

UNIT - I

Fundamentals of Image Processing and Image Transforms

asic ste s of Image Processing System Sam ling and Quanti ation of an image asic

relations i bet een i els Image ransforms: 2 D- Discrete ourier ransform, Discrete Cosine ransform DC ,avelet ransforms: Continuous avelet ransform, Discrete avelet ransforms.

UNIT - II 

Image Processing Techniques – Image Enhancement Spatial domain methods:

Histogram rocessing, undamentals of S atial filtering, Smoot ing s atial filters,

S ar ening s atial filters. re uency domain met ods: asics of filtering in fre uency domain, image smoot ing, image s ar ening, Selective filtering.

Image Segmentation

Segmentation conce ts, Point, ine and Edge Detection, res olding, Region ased

segmentation.

UNIT - III 

Image Compression

Image com ression fundamentals - Coding Redundancy, S atial and em oral redundancy,

Com ression models: ossy ossless, Huffman coding, rit metic coding, coding,

Run lengt coding, it lane coding, ransform coding, Predictive coding, avelet coding, PE Standards.

UNIT - IV

Basic steps of Video Processing

nalog Video, Digital Video. ime-Varying Image ormation models: ree-Dimensional

Motion Models, eometric Image ormation, P otometric Image ormation, Sam ling of Video signals, iltering o erations.

UNIT - V

2-D Motion Estimation

O tical flo , eneral Met odologies, Pi el ased Motion Esimation, loc - Matc ing

lgorit m, Mes based Motion Estimation, lobal Motion Estimation, Region based Motion

Estimation, Multi resolution motion estimation, aveform based coding, loc based

transform coding, Predictive coding, lication of motion estimation in Video coding.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 24

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

TEXT BOOKS

rd

1. Digital Image Processing on ale e and oods, 3 ed., Pearson.

2. Video rocessing and communication ao ang, oemOstermann and

st

a uin ang. 1 Ed., PH Int.

REFRENCES

1. Digital Video Processing M. e al, Prentice Hall International

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 25

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

MEC1061: RADAR SIGNAL PROCESSING

Lecture : 4 Periods/week Internal Marks : 40

External Marks : 60

Credits : 4 External Examination : 3 Hrs

-------------------------------------------------------------------------------------------------------------- 

UNIT - I 

Introduction Radar loc Diagram, Radar E uation, Information vailable from Radar

Ec o. Revie of Radar Range Performance eneral Radar Range E uation, Radar

Detection it oise amming, eacon and Re eater E uations, istatic Radar.Matc ed

ilter Receiver Im ulse Res onse, re uency Res onse C aracteristic and its Derivation, Matc ed ilter and Correlation unction, Correlation Detection and Cross-Correlation

Receiver. Efficiency of on-Matc ed ilters, Matc ed ilter for on- ite oise.

UNIT - II 

Detection of Radar Signals in oise: Detection Criteria eyman-Pearson Observer,

i eli ood-Ratio Receiver, Inverse Probability Receiver, Se uential Observer. Detectors

Envelo e Detector, ogarit mic Detector, I/Q Detector. utomatic Detection -C R

Receiver, Cell veraging C R Receiver, C R oss, C R ses in Radar. Radar Signal Management Sc ematics, Com onent Parts, Resources and Constraints.

UNIT - III 

aveform Selection 3, 2 : Radar mbiguity unction and mbiguity Diagram Princi les

and Pro erties S ecific Cases Ideal Case, Single Pulse of Sine ave, Periodic Pulse

rain, Single inear M Pulse, oiseli e aveforms. aveform Design

Re uirements.O timum aveforms for Detection in Clutter, amily of Radar aveforms.

UNIT - IV

Pulse Com ression in Radar Signals: Introduction, Significance, y es. inear M Pulse

Com ression loc Diagram, C aracteristics, Reduction of ime Sidelobes, Stretc

ec ni ues, eneration and Decoding of M aveforms loc Sc ematic and

C aracteristics of Passive System, Digital Com ression, S Pulse Com ression.

UNIT - V

P ase Coding ec ni ues: Princi les, inary P ase Coding, ar er Codes, Ma imal engt Se uences M S/ RS/P , loc Diagram of a P ase Coded C Radar. Poly P ase Codes: ran Codes, Costas Codes, on- inear M Pulse Com ression, Do ler olerant PC aveforms S ort Pulse, inear Period Modulation PM/H M . Sidelobe Reduction for P ase Coded PC Signals.

TEXT BOOKS

nd

1. Radar Handboo - M.I. S olni, 2 ed., 1991, Mc ra Hill.

2. Radar Design Princi les : Signal Processing and e Environment - red E.

nd

at anson,2 ed., 1999, PHI.

rd

3. Introduction to Radar Systems - M.I. S olni, 3 ed., 2001, MH.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 26

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

REFERENCES

1. Radar Princi les - Peyton . Peebles, r., 2004, o n iley.

2. Radar Signal Processing and da tive Systems - R. it berg, 1999, rtec House.

st

3. Radar Design Princi les - .E. at anson, 1 ed., 1969, Mc ra Hill.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 27

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

MEC1062: BIO-MEDICAL SIGNAL PROCESSING

Lecture : 4 Periods/week Internal Marks : 40

External Marks : 60

Credits : 4 External Examination : 3 Hrs

-------------------------------------------------------------------------------------------------------------- 

UNIT - I

Discrete and continuous Random variables, Probability distribution and density functions. aussian and Rayleig density functions, Correlation bet een random variables.

Stationary random rocess, Ergodicity, Po er s ectral density and autocorrelation function of random rocesses. oise o er s ectral density analysis, oise band idt , noise figure of systems.

UNIT- II

Data Compression Techniques: ossy and ossless data reduction lgorit ms. EC data

com ression using urning oint, EC, COR ES, Huffman coding, vector uantisation,

DC and t e transform.

UNIT- III 

Cardiological Signal Processing: Pre- rocessing. QRS Detection Met ods.R yt m

analysis. rr yt mia Detection lgorit ms. utomated EC nalysis.EC Pattern

Recognition.Heart rate variability analysis. da tive oise Cancelling: Princi les of da tive oise Cancelling. da tive oise Cancelling it t e MS da tation lgorit m. oise Cancelling Met od to En ance EC Monitoring. etal EC Monitoring.

UNIT- IV

Signal Averaging, polishing mean and trend removal, Prony s met od, Prony s Met od

based on t e east S uares Estimate, inear rediction. ule al er e uations,

nalysis of Evo ed Potentials.

UNIT- V

Neurological Signal Processing: Modeling of EE Signals. Detection of s i es and s indles Detection of l a, eta and amma aves. uto Regressive .R. modeling of sei ure EE . Slee Stage analysis. Inverse iltering. east s uares and olynomial modeling.

TEXT BOOKS

t

1. Probability, Random Variables Random Signal Princi les Peyton . Peebles, 4 ed., 2009, MH.

2. iomedical Signal Processing- Princi les and ec ni ues - D.C.Reddy, 2005, MH.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 28

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

REFERENCES

1. Digital io signal Processing - eit unat R, 1991, Elsevier.

2. iomedical Signal Processing - ay M, IEEE Press.

3. iomedical Signal Processing -Vol. I ime re uency nalysis - Co en. , 1986, CRC Press.

4. iomedical digital Signal Processing : C- anguage E eriments and aboratory E eriments, illis . om ins, PHI.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 29

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

MEC201: ADAPTIVE SIGNAL PROCESSING

Lecture : 4 Periods/week Internal Marks : 40

External Marks : 60

Credits : 4 External Examination : 3 Hrs

UNIT – I

Introduction to Adaptive Systems Adaptive Systems: Definitions, C aracteristics,

lications, E am le of an da tive System. e da tive inear Combiner - Descri tion,

eig t Vectors, Desired Res onse Performance function - radient Mean S uare Error.

UNIT – II

Development of Adaptive Filter Theory & Searching the Performance surface:

Introduction to iltering - Smoot ing and Prediction inear O timum iltering, Problem

statement, Princi le of Ort ogonality - Minimum Mean S uare Error, iener- Ho f

e uations, Error Performance - Minimum Mean S uare Error.

Searching the performance surface Met ods Ideas of radient Searc met ods -

radient Searc ing lgorit m its Solution - Stability Rate of convergence - earning

Curves.

UNIT - III 

Steepest Descent Algorithms

radient Searc by e tons Met od, Met od of Stee est Descent, Com arison of

earning Curves.

UNIT – IV

LMS Algorithm & Applications

Overvie - MS da tation algorit ms, Stability Performance analysis of MS lgorit ms -

MS radient Stoc astic algorit ms - Convergence of MS algorit m. Applications:

oise cancellation Cancellation of Ec oes in long distance tele one circuits, da tive eam forming.

UNIT – V

Kalman filtering:

Introduction - Recursive Mean S uare Estimation Random variables, Statement of alman

filtering roblem iltering -Initial conditions - Variants of alman filtering E tend alman

filtering.

TEXT BOOKS

1. da tive Signal Processing - ernard idro , Samuel D.Strearns, 2005, PE.

2. da tive ilter eory - Simon Hay in-, 4 ed., 2002,PE sia.

REFERENCES 

1. O timum signal rocessing: n introduction - So ocles. .Orfamadis, 2 ed., 1988, Mc ra -Hill, e yor .

2. da tive signal rocessing- eory and lications, S. omas le ander, 1986,

S ringer Verlag.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 30

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

MEC202: SPEECH PROCESSING

Lecture : 4 Periods/week Internal Marks : 40

External Marks : 60

Credits : 4 External Examination : 3 Hrs

-------------------------------------------------------------------------------------------------------------- 

UNIT - I

Fundamentals of Digital Speech Processing: natomy P ysiology of S eec Organs,

e rocess of S eec Production, e coustic eory of S eec Production, Digital

models fors eec signals.

UNIT - II 

Time Domain Models for Speech Processing

Introduction- indo considerations, S ort time energy and average magnitude S ort time

average ero crossing rate ,S eec vs. silence discrimination using energy and ero

crossing, Pitc eriod estimation using a arallel rocessing a roac , e s ort time

autocorrelation function, e s ort time average magnitude difference function, Pitc eriod

estimation using t e autocorrelation function.

UNIT – III 

Linear predictive coding (LPC) analysis

asic rinci les of inear Predictive nalysis: e utocorrelation Met od, e Covariance

Met od,Solution of c E uations: C oles y Decom osition Solution for Covariance

Met od, Durbin s Recursive Solution for t e utoCorrelation E uations, Com arision

bet een t e Met ods of Solution of t e PC nalysis E uations, lications of PC

Parameters: Pitc Detection using PC Parameters, ormant nalysis using PC Parameters.

Homomorphic Speech Processing

Introduction, Homomor ic Systems for Convolution: Pro erties of t e Com le Ce strum, Com utational Considerations, e Com le Ce strum of S eec , Pitc Detection, ormant Estimation, e Homomor icVocoder.

UNIT - IV

Speech enhancement: - ature of interfering sounds, S eec en ancment tec ni ues:

Single Micro one roac : s ectral substraction, En ancement by re-synt esis, Comb

filter, iener filter, Multimicro one roac .

Automatic speech recognition- asic attern recognition a roac es, Parametric

re resention of s eec , Evaluating t e similarity of s eec atterns, Isolated digit

Recognition System,.Contineous digit Recognition System

UNIT - V

Hidden Markov Model (HMM) for Speech

Hidden mar ov model HMM for s eec recognition, Viterbialgorit m, raining and testing

using HMMS, da ting to variability in s eec D , anguage models.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 31

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

Speaker recognition

Recognition tec ni ues, eatures t at distinguis s ea ers, S ea er Recognition Systems:

S ea er Verification System , S ea er Identification System.

TEXT BOOKS

1. Digital rocessing of s eec signals - .R Rabiner and S. .Sc afer. Pearson Education.

nd

2. S eec Communications: Human Mac ine - Douglas O S aug nessy, 2 ed., IEEE Press.

3. Digital rocessing of s eec signals. .R Rabinar and R Sc afer,1978, PHI.

REFERENCES

1. Discrete ime S eec Signal Processing : rinci les and Practice - omas . Quateri 1 ed., PE.

2. S eec udio Signal Processing- en old elson Morgan, 1 ed., iley.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 32

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

MEC203: SOC ARCHITECTURE

Lecture : 4 Periods/week Internal Marks : 40

External Marks : 60

Credits : 4 External Examination : 3 Hrs

-------------------------------------------------------------------------------------------------------------- 

UNIT - I

Introduction to Processor Design:

bstraction in Hard are Design, M O a sim le rocessor , Processor design trade off,

Design for lo o er consum tion.

ARM Processor as System-on-Chip: corn RISC Mac ine rc itecture in eritance

RM rogramming model RM develo ment tools 3 and 5 stage i eline RM

organi ation RM instruction e ecution and im lementation RM Co- rocessor interface

UNIT - II 

ARM Assembly Language Programming:

RM instruction ty es data transfer, data rocessing and control flo instructions RM

instruction set Co- rocessor instructions.

Architectural Support for High Level Language: Data ty es abstraction in Soft are

design E ressions oo s unctions and Procedures Conditional Statements se

of Memory

UNIT - III 

Memory Hierarchy: Memory si e and s eed On-c i memory Cac es Cac e design-an e am le memory management

UNIT - IV

Architectural Support for System Development: dvanced Microcontroller bus

arc itecture RM memory interface RM reference eri eral s ecification Hard are

system rototy ing tools rmulator Debug arc itecture

UNIT - V

Architectural Support for Operating System: n introduction to O erating Systems

RM system control co rocessor CP15 rotection unit registers RM rotection unit

CP15 MM registers RM MM rc itecture Sync roni ation Conte t S itc ing in ut

and out ut

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 33

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

TEXT BOOKS

nd

1. RM System on C i rc itecture Steve urber 2 ed., 2000, ddison esley

Professional.

st

2. Design of System on a C i : Devices and Com onents Ricardo Reis, 1 ed., 2004, S ringer

REFERENCES 

1. Co-Verification of Hard are and Soft are for RM System on C i Design

Embedded ec nology ason ndre s e nes, and CDROM System on

C i Verification Met odologies and ec ni ues Pra as Ras in ar, Peter

Paterson and eena Sing , 2001, lu er cademic Publis ers.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 34

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

MEC204 : CODING THEORY AND TECHNIQUES

Lecture : 4 Periods/week Internal Marks : 40

External Marks : 60

Credits : 4 External Examination : 3 Hrs

-------------------------------------------------------------------------------------------------------------- 

UNIT - I

Coding for Reliable Digital Transmission and storage: Mat ematical model of

Information, ogarit mic Measure of Information, verage and Mutual Information and

Entro y, y es of Errors, Error Control Strategies. inear loc Codes: Introduction to

inear loc Codes, Syndrome and Error Detection, Minimum Distance of a loc code, Error-Detecting and Error-correcting Ca abilities of a loc code, Standard array and Syndrome Decoding, Probability of an undetected error for inear Codes over a SC,

Hamming Codes. lications of loc codes for Error control in data storage system

UNIT- II

Cyclic codes: Descri tion, enerator and Parity-c ec Matrices, Encoding, Syndrome Com utation and Error Detection, Decoding ,Cyclic Hamming Codes, S ortened cyclic codes, Error-tra ing decoding for cyclic codes, Ma ority logic decoding for cyclic codes.

UNIT- III 

Convolutional codes: Encoding of Convolutional Codes, Structural and Distance Pro erties, ma imum li eli ood decoding, Se uential decoding, Ma ority- logic decoding of

Convolution codes. lication of Viterbi Decoding and Se uential Decoding, lications of

Convolutional codes in RQ system.

UNIT- IV

urst Error-Correcting codes: Decoding of Signle- urst error Correcting Cyclic codes, Single- urst-Error-Correcting Cyclic codes, urst-Error-Correcting Convoulutional Codes,

ounds on urst Error-Correcting Ca ability, Interleaved Cyclic and Convolutional Codes , P ased- urst Error-Correcting Cyclic and Convolutional codes.

UNIT – V

BCH – Codes: CH code- Definition, Minimum distance and CH ounds, Decoding Procedure for CH Codes- Syndrome Com utation and Iterative lgorit ms, Error ocation Polynomials and umbers for single and double error correction

TEXT BOOKS

1. Error Control Coding- undamentals and lications S u in, Daniel .Costello, r,

Prentice Hall, Inc.

2. Error Correcting Coding eory-Man oung R ee- 1989, Mc ra -Hill Publis ing.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 35

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 36

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

MEC2051 : CPLD & FPGA ARCHITECTURES AND APPLICATIONS

Lecture : 4 Periods/week Internal Marks : 40

External Marks : 60

Credits : 4 External Examination : 3 Hrs

-------------------------------------------------------------------------------------------------------------- 

UNIT - I

Programmable logic : ROM, P , P P D, P eatures, rogramming and

a lications using com le rogrammable logic devices ltera series Ma 5000/7000

series and ltera E logic-10000 series CP D, MD s- CP D Mac 1to 5 , Cy res

SH 370 Device tec nology, attice P S s arc itectures 3000 series S eed

erformance and in system rogrammability.

UNIT - II 

FPGAs: ield Programmable gate arrays- ogic bloc s, routing arc itecture, design flo

tec nology ma ing for P s, Case studies itir C4000 ER s E

8000/10000 P s: ORC s O timi ed Reconfigurable Cell rray : C E s C -

1,2,3 and t eir s eed erformance

UNIT - III 

lternative reali ation for state mac ine c at suing micro rogramming lin ed state mac ine

one ot state mac ine, etrinetes for state mac ines-basic conce ts, ro erties, e tended

etrinetes for arallel controllers.

UNIT - IV

Digital front end digital design tools for P s SICs: sing mentor gra ics ED tool

P dvantage Design flo using P s

UNIT - V

Case studies of araller adder cell araller adder se uential circuits, counters, multi le ers, arellel controllers.

TEXT BOOKS

1. ield Programmable ate rray ec nology - S. rimberger, Edr, 1994, lu er cademic Publications.

2. ield Programmable ate rrays, o n V.Oldfield, Ric ard C Dore, iley

Publications.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 37

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

REFERENCES

1. Digital Design sing ield Programmable ate rray, P. .C an S. Mourad, 1994, Prentice Hall.

2. Digital System Design using Programmable ogic Devices Parag. . ala, 2003, SP.

3. ield rogrammable gate array, S. ro n, R. . rancis, .Rose, . .Vranesic, 2007, SP.

4. Digital Systems Design it P s and CP Ds Ian rout, 2009, Elsevier

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 38

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

MEC2052 : DESIGN FOR TESTABILITY

Lecture : 4 Periods/week Internal Marks : 40

External Marks : 60

Credits : 4 External Examination : 3 Hrs

UNIT – I

Introduction to Test and Design for Testability (DFT) Fundamentals

Modeling: Modeling Digital Circuits at ogic evel, register evel, and Structural Models. evels of Modeling. ogic Simulation: y es of Simulation, Delay Models, Element Evaluation, Ha ard Detection, ate evel Event Driven Simulation.

UNIT – II

Fault Modeling: ogic ault Models, ault Detection and Redundancy, ault e uivalence and ault ocation. Single Stuc and Multi le Stuc - ault Models, ault Simulation lications, eneral ec ni ues for Combinational Circuits.

UNIT – III 

esting for Single Stuc aults SS utomated est Pattern eneration P / for

SS s in Combinational and Se uential Circuits, unctional esting it S ecific ault

Models, Vector Simulation P Vectors, ormats, Com action and Com ression,

Selecting P ool.

UNIT – IV

Design for estability testability rade-off s ec ni ues, Scan rc itectures and esting, Controllability and bsorbability, eneric oundary Scan, ull Integrated Scan, Storage Cells foe Scan Design, oard level and System level a roac es, oundary Scans Standards, Com ression ec ni ues Different ec ni ues, Syndrome test and Signature analysis.

UNIT – V

uilt-in Self test IS IS Conce ts and est attern eneration. S ecific IS

rc itectures OCS , S MPS, C IS , R D, I O. rief ideas on some advanced

IS conce ts and design for self-test at board level. Memory IS M IS : Memory est rc itectures and ec ni ues, Introduction to Memory est, y es of Memories and Integration, Embedded Memory esting Model, Memory est re uirements for M IS , esting eatures.

TEXT BOOKS

1. Digital Systems esting and estable Design Miron bramovici, Melvin . reur, rt uD. riedman, o n iley Sons.

2. Design for est for Digital ICs Embedded Core Systems lfred Crouc , 2008,

PE.

3. Introduction to V SI esting Robrt. . eugate , Steven M.Mclntyre, Engle ood

Cliffs, 1988, Prentice Hall.

REFERENCES 

1. Essentials of Electronic esting M. . us nell, Vis ani.D. gar al, S ringer.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 39

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

MEC2061 : WIRELESS COMMUNICATION AND NETWORKS

Lecture : 4 Periods/week Internal Marks : 40

External Marks : 60

Credits : 4 External Examination : 3 Hrs

UNIT - I 

Wireless Communications & System Fundamentals: Introduction to ireless

communications systems, e am les, com arisons trends, Cellular conce ts-fre uency

reuse, strategies, interference system ca acity, truc ing grade of service, im roving

coverage ca acity in cellular systems.

UNIT - II 

Multiple Access Techniques for Wireless Communication: DM , DM , SSM

HM /CDM /Hybrid tec ni ues , SDM tec ni ue S a licable to ireless

communications .Pac et radio access- rotocols, CSM rotocols, reservation rotocols,

ca ture effect in ac et radio, ca acity of cellular systems.

UNIT - III 

Wireless Networking: Introduction, differences in ireless fi ed tele one net or s,

traffic routing in ireless net or s circuit s itc ing, ac et s itc ing .25 rotocol.

Wireless data services cellular digital ac et data CDPD , advanced radio data

information systems, R M mobile data RMD . Common c annel signaling CCS , ISD -

road band ISD M, Signaling System no .7 SS7 - rotocols, net or services art,

user art, signaling traffic, services erformance

UNIT - IV

Mobile IP and Wireless Application Protocol: Mobile IP O eration of mobile IP, Co 

located address, Registration, unneling, P rc itecture, overvie , M scri ts, P

service, P session rotocol, ireless transaction, ireless datagram rotocol. ireless

ec nology, Infrared s, S read s ectrum s, arro ban micro ave s,

IEEE 802 rotocol rc itecture, IEEE802 arc itecture and services, 802.11 medium access

control, 802.11 ysical layer.

UNIT - V

Mobile Data Networks: Introduction, Data oriented CDPD et or , PRS and ig er data

rates, S ort messaging service in SM, Mobile a lication rotocol. d- oc ireless

et or s: Cellular and d oc ireless net or s, a lications, M C rotocols, Routing,

Multicasting, rans ort layer Protocols, uality of service bro sing, de loyment

considerations, d oc ireless Internet

TEXT BOOKS

1. ireless Communication and et or ing illiam Stallings, 2003, PHI.

nd

2. ireless Communications, Princi les, Practice- eodore, S.Ra a ort, 2 Ed. 2002, PHI.

3. Princi les of ireless et or s ave Pa aven and P. ris na Murt y, 2002, Pearson Education ublis ers

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 40

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

REFERENCES

1. ireless Digital Communications amilo e er, 1999, PHI.

2. elecommunication System Engineering Roger .reeman, 4/ed., iley-Interscience, o n iley Sons, 2004.

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 41

 

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING (AUTONOMOUS), MYLAVARAM - 521230

MEC2062 : VLSI SIGNAL PROCESSING

Lecture : 4 Periods/week Internal Marks : 40

External Marks : 60

Credits : 4 External Examination : 3 Hrs

UNIT - I

Introduction to DSP: y ical DSP algorit ms, DSP algorit ms benefits, Re resentation of DSP algorit ms Pipelining and Parallel Processing: Introduction, Pi elining of IR Digital filters, Parallel Processing, Pi elining and Parallel Processing for o Po er Retiming:

Introduction Definitions and Pro erties Solving System of Ine ualities Retiming

ec ni ues

UNIT - II 

Folding and Unfolding: olding : Introduction - olding ransform - Register minimi ation ec ni ues Register minimi ation in folded arc itectures folding of multirate systems

Unfolding: Introduction n lgorit m for nfolding Pro erties of nfolding critical

Pat , nfolding and Retiming lications of nfolding

UNIT - III 

Systolic Architecture Design: Introduction Systolic rray Design Met odology IR

Systolic rrays Selection of Sc eduling Vector Matri Multi lication and 2D Systolic

rray Design Systolic Design for S ace Re resentations contain Delays

UNIT – IV

Fast Convolution: Introduction Coo - oom lgorit m inogard algorit m Iterated

Convolution Cyclic Convolution Design of ast Convolution algorit m by Ins ection

UNIT – V

Low Power Design: Scaling Vs Po er Consum tion Po er nalysis, Po er Reduction

tec ni ues Po er Estimation roac es Programmable DSP : Evaluation of

Programmable Digital Signal Processors, DSP Processors for Mobile and ireless

Communications, Processors for Multimedia Signal Processing

TEXT BOOKS

1. V SI Digital Signal Processing- System Design and Im lementation es ab .

Part i, 1998, iley Inter Science.

2. V SI and Modern Signal rocessing ung S. , H. .ile House, . ailat ,

1985, Prentice Hall.

REFERENCES

1. Design of nalog Digital V SI Circuits for elecommunications and Signal

Processing ose E. rance, annis sividis, 1994, Prentice Hall.

2. V SI Digital Signal Processing Medisetti V. ,1995, IEEE Press , S

M.TECH (SYSTEMS AND SIGNAL PROCESSING), A.Y.2010-2011 Page 42