Improving sampling, optimization and feature extraction in Boltzmann machines

L’apprentissage supervisé de réseaux hiérarchiques à grande échelle connaît présentement un succès fulgurant. Malgré cette effervescence, l’apprentissage non-supervisé représente toujours, selon plusieurs chercheurs, un élément clé de l’Intelligence Artificielle, où les agents doivent apprendre à partir d’un nombre potentiellement limité de données. Cette thèse s’inscrit dans cette pensée et aborde divers sujets de recherche liés au problème d’estimation de densité par l’entremise des machines de Boltzmann (BM), modèles graphiques probabilistes au coeur de l’apprentissage profond. Nos contributions touchent les domaines de l’échantillonnage, l’estimation de fonctions de partition, l’optimisation ainsi que l’apprentissage de représentations invariantes. Cette thèse débute par l’exposition d’un nouvel algorithme d'échantillonnage adaptatif, qui ajuste (de fa ̧con automatique) la température des chaînes de Markov sous simulation, afin de maintenir une vitesse de convergence élevée tout au long de l’apprentissage. Lorsqu’utilisé dans le contexte de l’apprentissage par maximum de vraisemblance stochastique (SML), notre algorithme engendre une robustesse accrue face à la sélection du taux d’apprentissage, ainsi qu’une meilleure vitesse de convergence. Nos résultats sont présent ́es dans le domaine des BMs, mais la méthode est générale et applicable à l’apprentissage de tout modèle probabiliste exploitant l’échantillonnage par chaînes de Markov. Tandis que le gradient du maximum de vraisemblance peut-être approximé par échantillonnage, l’évaluation de la log-vraisemblance nécessite un estimé de la fonction de partition. Contrairement aux approches traditionnelles qui considèrent un modèle donné comme une boîte noire, nous proposons plutôt d’exploiter la dynamique de l’apprentissage en estimant les changements successifs de log-partition encourus à chaque mise à jour des paramètres. Le problème d’estimation est reformulé comme un problème d’inférence similaire au filtre de Kalman, mais sur un graphe bi-dimensionnel, où les dimensions correspondent aux axes du temps et au paramètre de température. Sur le thème de l’optimisation, nous présentons également un algorithme permettant d’appliquer, de manière efficace, le gradient naturel à des machines de Boltzmann comportant des milliers d’unités. Jusqu’à présent, son adoption était limitée par son haut coût computationel ainsi que sa demande en mémoire. Notre algorithme, Metric-Free Natural Gradient (MFNG), permet d’éviter le calcul explicite de la matrice d’information de Fisher (et son inverse) en exploitant un solveur linéaire combiné à un produit matrice-vecteur efficace. L’algorithme est prometteur: en terme du nombre d’évaluations de fonctions, MFNG converge plus rapidement que SML. Son implémentation demeure malheureusement inefficace en temps de calcul. Ces travaux explorent également les mécanismes sous-jacents à l’apprentissage de représentations invariantes. À cette fin, nous utilisons la famille de machines de Boltzmann restreintes “spike & slab” (ssRBM), que nous modifions afin de pouvoir modéliser des distributions binaires et parcimonieuses. Les variables latentes binaires de la ssRBM peuvent être rendues invariantes à un sous-espace vectoriel, en associant à chacune d’elles, un vecteur de variables latentes continues (dénommées “slabs”). Ceci se traduit par une invariance accrue au niveau de la représentation et un meilleur taux de classification lorsque peu de données étiquetées sont disponibles. Nous terminons cette thèse sur un sujet ambitieux: l’apprentissage de représentations pouvant séparer les facteurs de variations présents dans le signal d’entrée. Nous proposons une solution à base de ssRBM bilinéaire (avec deux groupes de facteurs latents) et formulons le problème comme l’un de “pooling” dans des sous-espaces vectoriels complémentaires.

Analysis of Terahertz Detection with A 2D Hot-Electron Quantum Well Detector

The operation of a previously proposed terahertz (THZ) detector is formulated in detail. The detector is based on the hot-electron effect of the 2D electron gas (2DEG) in the quantum well (QW) of a GaAs/AIGaAs heterostructure. The interaction between the THz radiation and the 2DEG, the current enhancement due to hot -electron effect, and the noise performance of the detector are analyzed

“Improved Feature Extraction and Classification Techniques for Multispectral Brain Magnetic Resonance Images”

Magnetic Resonance Imaging (MRI) is a multi sequence medical imaging technique in which stacks of images are acquired with different tissue contrasts. Simultaneous observation and quantitative analysis of normal brain tissues and small abnormalities from these large numbers of different sequences is a great challenge in clinical applications. Multispectral MRI analysis can simplify the job considerably by combining unlimited number of available co-registered sequences in a single suite. However, poor performance of the multispectral system with conventional image classification and segmentation methods makes it inappropriate for clinical analysis. Recent works in multispectral brain MRI analysis attempted to resolve this issue by improved feature extraction approaches, such as transform based methods, fuzzy approaches, algebraic techniques and so forth. Transform based feature extraction methods like Independent Component Analysis (ICA) and its extensions have been effectively used in recent studies to improve the performance of multispectral brain MRI analysis. However, these global transforms were found to be inefficient and inconsistent in identifying less frequently occurred features like small lesions, from large amount of MR data. The present thesis focuses on the improvement in ICA based feature extraction techniques to enhance the performance of multispectral brain MRI analysis. Methods using spectral clustering and wavelet transforms are proposed to resolve the inefficiency of ICA in identifying small abnormalities, and problems due to ICA over-completeness. Effectiveness of the new methods in brain tissue classification and segmentation is confirmed by a detailed quantitative and qualitative analysis with synthetic and clinical, normal and abnormal, data. In comparison to conventional classification techniques, proposed algorithms provide better performance in classification of normal brain tissues and significant small abnormalities.

Feature Extraction Methods Based on Linear Predictive Coding and Wavelet Packet Decomposition for Recognizing Spoken Words in Malayalam

Speech signals are one of the most important means of communication among the human beings. In this paper, a comparative study of two feature extraction techniques are carried out for recognizing speaker independent spoken isolated words. First one is a hybrid approach with Linear Predictive Coding (LPC) and Artificial Neural Networks (ANN) and the second method uses a combination of Wavelet Packet Decomposition (WPD) and Artificial Neural Networks. Voice signals are sampled directly from the microphone and then they are processed using these two techniques for extracting the features. Words from Malayalam, one of the four major Dravidian languages of southern India are chosen for recognition. Training, testing and pattern recognition are performed using Artificial Neural Networks. Back propagation method is used to train the ANN. The proposed method is implemented for 50 speakers uttering 20 isolated words each. Both the methods produce good recognition accuracy. But Wavelet Packet Decomposition is found to be more suitable for recognizing speech because of its multi-resolution characteristics and efficient time frequency localizations

Effectiveness Of Feature Detection Operators On The Performance Of Iris Biometric Recognition System

Iris Recognition is a highly efficient biometric identification system with great possibilities for future in the security systems area.Its robustness and unobtrusiveness, as opposed tomost of the currently deployed systems, make it a good candidate to replace most of thesecurity systems around. By making use of the distinctiveness of iris patterns, iris recognition systems obtain a unique mapping for each person. Identification of this person is possible by applying appropriate matching algorithm.In this paper, Daugman’s Rubber Sheet model is employed for irisnormalization and unwrapping, descriptive statistical analysis of different feature detection operators is performed, features extracted is encoded using Haar wavelets and for classification hammingdistance as a matching algorithm is used. The system was tested on the UBIRIS database. The edge detection algorithm, Canny, is found to be the best one to extract most of the iris texture. The success rate of feature detection using canny is 81%, False Accept Rate is 9% and False Reject Rate is 10%.

COMBINED FEATURE EXTRACTION TECHNIQUES AND NAIVE BAYES CLASSIFIER FOR SPEECH RECOGNITION

Speech processing and consequent recognition are important areas of Digital Signal Processing since speech allows people to communicate more natu-rally and efficiently. In this work, a speech recognition system is developed for re-cognizing digits in Malayalam. For recognizing speech, features are to be ex-tracted from speech and hence feature extraction method plays an important role in speech recognition. Here, front end processing for extracting the features is per-formed using two wavelet based methods namely Discrete Wavelet Transforms (DWT) and Wavelet Packet Decomposition (WPD). Naive Bayes classifier is used for classification purpose. After classification using Naive Bayes classifier, DWT produced a recognition accuracy of 83.5% and WPD produced an accuracy of 80.7%. This paper is intended to devise a new feature extraction method which produces improvements in the recognition accuracy. So, a new method called Dis-crete Wavelet Packet Decomposition (DWPD) is introduced which utilizes the hy-brid features of both DWT and WPD. The performance of this new approach is evaluated and it produced an improved recognition accuracy of 86.2% along with Naive Bayes classifier.

IRIS BIOMETRIC RECOGNITION SYSTEM EMPLOYING CANNY OPERATOR

Biometrics has become important in security applications. In comparison with many other biometric features, iris recognition has very high recognition accuracy because it depends on iris which is located in a place that still stable throughout human life and the probability to find two identical iris's is close to zero. The identification system consists of several stages including segmentation stage which is the most serious and critical one. The current segmentation methods still have limitation in localizing the iris due to circular shape consideration of the pupil. In this research, Daugman method is done to investigate the segmentation techniques. Eyelid detection is another step that has been included in this study as a part of segmentation stage to localize the iris accurately and remove unwanted area that might be included. The obtained iris region is encoded using haar wavelets to construct the iris code, which contains the most discriminating feature in the iris pattern. Hamming distance is used for comparison of iris templates in the recognition stage. The dataset which is used for the study is UBIRIS database. A comparative study of different edge detector operator is performed. It is observed that canny operator is best suited to extract most of the edges to generate the iris code for comparison. Recognition rate of 89% and rejection rate of 95% is achieved

A Comparative Study of Wavelet Based Feature Extraction Techniques in Recognizing Isolated Spoken Words

Speech is a natural mode of communication for people and speech recognition is an intensive area of research due to its versatile applications. This paper presents a comparative study of various feature extraction methods based on wavelets for recognizing isolated spoken words. Isolated words from Malayalam, one of the four major Dravidian languages of southern India are chosen for recognition. This work includes two speech recognition methods. First one is a hybrid approach with Discrete Wavelet Transforms and Artificial Neural Networks and the second method uses a combination of Wavelet Packet Decomposition and Artificial Neural Networks. Features are extracted by using Discrete Wavelet Transforms (DWT) and Wavelet Packet Decomposition (WPD). Training, testing and pattern recognition are performed using Artificial Neural Networks (ANN). The proposed method is implemented for 50 speakers uttering 20 isolated words each. The experimental results obtained show the efficiency of these techniques in recognizing speech

FEATURE SELECTION AND COMPARISON OF TWO NAÏVE BAYES CLASSIFICATION METHODS IN THE CONTEXT OF SPAM FILTERING

Treating e-mail filtering as a binary text classification problem, researchers have applied several statistical learning algorithms to email corpora with promising results. This paper examines the performance of a Naive Bayes classifier using different approaches to feature selection and tokenization on different email corpora

Multifinger Feature Level Fusion Based Fingerprint Identification

Fingerprint based authentication systems are one of the cost-effective biometric authentication techniques employed for personal identification. As the data base population increases, fast identification/recognition algorithms are required with high accuracy. Accuracy can be increased using multimodal evidences collected by multiple biometric traits. In this work, consecutive fingerprint images are taken, global singularities are located using directional field strength and their local orientation vector is formulated with respect to the base line of the finger. Feature level fusion is carried out and a 32 element feature template is obtained. A matching score is formulated for the identification and 100% accuracy was obtained for a database of 300 persons. The polygonal feature vector helps to reduce the size of the feature database from the present 70-100 minutiae features to just 32 features and also a lower matching threshold can be fixed compared to single finger based identification

Development of selected electro-active polymer-ceramic nanocomposites as pyroelectric thermal/infrared detector materials

The present work emphasises on the synthesis and characterization of electro-active polymer-ceramic nanocomposites which can be used for pyroelectric thermal/infrared detection applications. Two sets of samples belong to polymer-microcrystalline composites have also been investigated in the work. The polymers used in the work have been commercially available ones, but the nanoceramics have been synthesized following simple chemical routes and aqueous organic gel routes. After characterizing the nanoceramics for their structure by powder XRD, they have been dispersed in liquid polymer and sonicated for uniform dispersion. The viscous mixture so formed was cast in the form of films for experimentation. Samples with volume fraction of the ceramic phase varied from 0 to 0.25 have been prepared. Solution growth was followed to prepare microcrystalline samples for the polymer-microcrystalline composites. The physical properties that determine the pyroelectric sensitivity of a material are dielectric constant, dielectric loss, pyroelectric coefficient, thermal conductivity and specific heat capacity. These parameters have been determined for all the samples and compositions reported in this work.The pyroelectric figures of merit for all the samples were determined. The pyroelectric figures of merit that determine the pyroelectric sensitivity of a material are current sensitivity, voltage responsivity and detectivity. All these have been determined for each set of samples and reported in the thesis. In order to assess the flexibility and mouldability of the composites we have measured the Shore hardness of each of the composites by indentation technique and compared with the pyroelectric figures of merit. Some important factors considered during the material fabrication stages were maximum flexibility and maximum figures of merit for pyroelectric thermal/IR detection applications. In order to achieve these goals, all the samples are synthesized as composites of polymers and nano/microcrystalline particles and are prepared in the form of freestanding films. The selected polymer matrices and particle inclusions possess good pyroelectric coefficients, low thermal and dielectric properties, so that good pyroelectric figures of merit could be achieved. The salient features of the work include the particle size of the selected ceramic materials. Since they are in nanometer size it was possible to achieve high flexibility and moldability with high figures of merit for even low volume fractions of inclusions of the prepared nanocrystalline composites. In the case of microcrystalline TGS and DTGS, their composites in PU matrix protect them from fragility and humidity susceptibility and made them for environmental friendly applications.

Region-Based Feature Interpretation for Recognizing 3D Models in 2D Images

In model-based vision, there are a huge number of possible ways to match model features to image features. In addition to model shape constraints, there are important match-independent constraints that can efficiently reduce the search without the combinatorics of matching. I demonstrate two specific modules in the context of a complete recognition system, Reggie. The first is a region-based grouping mechanism to find groups of image features that are likely to come from a single object. The second is an interpretive matching scheme to make explicit hypotheses about occlusion and instabilities in the image features.

Feature Point Detection and Curve Approximation for Early Processing of Freehand Sketches

Freehand sketching is both a natural and crucial part of design, yet is unsupported by current design automation software. We are working to combine the flexibility and ease of use of paper and pencil with the processing power of a computer to produce a design environment that feels as natural as paper, yet is considerably smarter. One of the most basic steps in accomplishing this is converting the original digitized pen strokes in the sketch into the intended geometric objects using feature point detection and approximation. We demonstrate how multiple sources of information can be combined for feature detection in strokes and apply this technique using two approaches to signal processing, one using simple average based thresholding and a second using scale space.

Feature Selection for Face Detection

We present a new method to select features for a face detection system using Support Vector Machines (SVMs). In the first step we reduce the dimensionality of the input space by projecting the data into a subset of eigenvectors. The dimension of the subset is determined by a classification criterion based on minimizing a bound on the expected error probability of an SVM. In the second step we select features from the SVM feature space by removing those that have low contributions to the decision function of the SVM.

On the difficulty of feature-based attentional modulations in visual object recognition: A modeling study.

Numerous psychophysical experiments have shown an important role for attentional modulations in vision. Behaviorally, allocation of attention can improve performance in object detection and recognition tasks. At the neural level, attention increases firing rates of neurons in visual cortex whose preferred stimulus is currently attended to. However, it is not yet known how these two phenomena are linked, i.e., how the visual system could be "tuned" in a task-dependent fashion to improve task performance. To answer this question, we performed simulations with the HMAX model of object recognition in cortex [45]. We modulated firing rates of model neurons in accordance with experimental results about effects of feature-based attention on single neurons and measured changes in the model's performance in a variety of object recognition tasks. It turned out that recognition performance could only be improved under very limited circumstances and that attentional influences on the process of object recognition per se tend to display a lack of specificity or raise false alarm rates. These observations lead us to postulate a new role for the observed attention-related neural response modulations.