Modeling High-Dimensional Audio Sequences with Recurrent Neural Networks

Cette thèse étudie des modèles de séquences de haute dimension basés sur des réseaux de neurones récurrents (RNN) et leur application à la musique et à la parole. Bien qu'en principe les RNN puissent représenter les dépendances à long terme et la dynamique temporelle complexe propres aux séquences d'intérêt comme la vidéo, l'audio et la langue naturelle, ceux-ci n'ont pas été utilisés à leur plein potentiel depuis leur introduction par Rumelhart et al. (1986a) en raison de la difficulté de les entraîner efficacement par descente de gradient. Récemment, l'application fructueuse de l'optimisation Hessian-free et d'autres techniques d'entraînement avancées ont entraîné la recrudescence de leur utilisation dans plusieurs systèmes de l'état de l'art. Le travail de cette thèse prend part à ce développement. L'idée centrale consiste à exploiter la flexibilité des RNN pour apprendre une description probabiliste de séquences de symboles, c'est-à-dire une information de haut niveau associée aux signaux observés, qui en retour pourra servir d'à priori pour améliorer la précision de la recherche d'information. Par exemple, en modélisant l'évolution de groupes de notes dans la musique polyphonique, d'accords dans une progression harmonique, de phonèmes dans un énoncé oral ou encore de sources individuelles dans un mélange audio, nous pouvons améliorer significativement les méthodes de transcription polyphonique, de reconnaissance d'accords, de reconnaissance de la parole et de séparation de sources audio respectivement. L'application pratique de nos modèles à ces tâches est détaillée dans les quatre derniers articles présentés dans cette thèse. Dans le premier article, nous remplaçons la couche de sortie d'un RNN par des machines de Boltzmann restreintes conditionnelles pour décrire des distributions de sortie multimodales beaucoup plus riches. Dans le deuxième article, nous évaluons et proposons des méthodes avancées pour entraîner les RNN. Dans les quatre derniers articles, nous examinons différentes façons de combiner nos modèles symboliques à des réseaux profonds et à la factorisation matricielle non-négative, notamment par des produits d'experts, des architectures entrée/sortie et des cadres génératifs généralisant les modèles de Markov cachés. Nous proposons et analysons également des méthodes d'inférence efficaces pour ces modèles, telles la recherche vorace chronologique, la recherche en faisceau à haute dimension, la recherche en faisceau élagué et la descente de gradient. Finalement, nous abordons les questions de l'étiquette biaisée, du maître imposant, du lissage temporel, de la régularisation et du pré-entraînement.

Leveraging noisy side information for disentangling of factors of variation in a supervised setting

Ce mémoire est composé de trois articles et présente les résultats de travaux de recherche effectués dans le but d'améliorer les techniques actuelles permettant d'utiliser des données associées à certaines tâches dans le but d'aider à l'entraînement de réseaux de neurones sur une tâche différente. Les deux premiers articles présentent de nouveaux ensembles de données créés pour permettre une meilleure évaluation de ce type de techniques d'apprentissage machine. Le premier article introduit une suite d'ensembles de données pour la tâche de reconnaissance automatique de chiffres écrits à la main. Ces ensembles de données ont été générés à partir d'un ensemble de données déjà existant, MNIST, auquel des nouveaux facteurs de variation ont été ajoutés. Le deuxième article introduit un ensemble de données pour la tâche de reconnaissance automatique d'expressions faciales. Cet ensemble de données est composé d'images de visages qui ont été collectées automatiquement à partir du Web et ensuite étiquetées. Le troisième et dernier article présente deux nouvelles approches, dans le contexte de l'apprentissage multi-tâches, pour tirer avantage de données pour une tâche donnée afin d'améliorer les performances d'un modèle sur une tâche différente. La première approche est une généralisation des neurones Maxout récemment proposées alors que la deuxième consiste en l'application dans un contexte supervisé d'une technique permettant d'inciter des neurones à apprendre des fonctions orthogonales, à l'origine proposée pour utilisation dans un contexte semi-supervisé.

Distributed conditional computation

L'objectif de cette thèse est de présenter différentes applications du programme de recherche de calcul conditionnel distribué. On espère que ces applications, ainsi que la théorie présentée ici, mènera à une solution générale du problème d'intelligence artificielle, en particulier en ce qui a trait à la nécessité d'efficience. La vision du calcul conditionnel distribué consiste à accélérer l'évaluation et l'entraînement de modèles profonds, ce qui est très différent de l'objectif usuel d'améliorer sa capacité de généralisation et d'optimisation. Le travail présenté ici a des liens étroits avec les modèles de type mélange d'experts. Dans le chapitre 2, nous présentons un nouvel algorithme d'apprentissage profond qui utilise une forme simple d'apprentissage par renforcement sur un modèle d'arbre de décisions à base de réseau de neurones. Nous démontrons la nécessité d'une contrainte d'équilibre pour maintenir la distribution d'exemples aux experts uniforme et empêcher les monopoles. Pour rendre le calcul efficient, l'entrainement et l'évaluation sont contraints à être éparse en utilisant un routeur échantillonnant des experts d'une distribution multinomiale étant donné un exemple. Dans le chapitre 3, nous présentons un nouveau modèle profond constitué d'une représentation éparse divisée en segments d'experts. Un modèle de langue à base de réseau de neurones est construit à partir des transformations éparses entre ces segments. L'opération éparse par bloc est implémentée pour utilisation sur des cartes graphiques. Sa vitesse est comparée à deux opérations denses du même calibre pour démontrer le gain réel de calcul qui peut être obtenu. Un modèle profond utilisant des opérations éparses contrôlées par un routeur distinct des experts est entraîné sur un ensemble de données d'un milliard de mots. Un nouvel algorithme de partitionnement de données est appliqué sur un ensemble de mots pour hiérarchiser la couche de sortie d'un modèle de langage, la rendant ainsi beaucoup plus efficiente. Le travail présenté dans cette thèse est au centre de la vision de calcul conditionnel distribué émis par Yoshua Bengio. Elle tente d'appliquer la recherche dans le domaine des mélanges d'experts aux modèles profonds pour améliorer leur vitesse ainsi que leur capacité d'optimisation. Nous croyons que la théorie et les expériences de cette thèse sont une étape importante sur la voie du calcul conditionnel distribué car elle cadre bien le problème, surtout en ce qui concerne la compétitivité des systèmes d'experts.

Neural changes associated with semantic processing in healthy aging despite intact behavioral performance

Semantic memory recruits an extensive neural network including the left inferior prefrontal cortex (IPC) and the left temporoparietal region, which are involved in semantic control processes, as well as the anterior temporal lobe region (ATL) which is considered to be involved in processing semantic information at a central level. However, little is known about the underlying neuronal integrity of the semantic network in normal aging. Young and older healthy adults carried out a semantic judgment task while their cortical activity was recorded using magnetoencephalography (MEG). Despite equivalent behavioral performance, young adults activated the left IPC to a greater extent than older adults, while the latter group recruited the temporoparietal region bilaterally and the left ATL to a greater extent than younger adults. Results indicate that significant neuronal changes occur in normal aging, mainly in regions underlying semantic control processes, despite an apparent stability in performance at the behavioral level.

Development and Evaluation of Blind Identification Techniques for Nonlinear Systems

Identification and Control of Non‐linear dynamical systems are challenging problems to the control engineers.The topic is equally relevant in communication,weather prediction ,bio medical systems and even in social systems,where nonlinearity is an integral part of the system behavior.Most of the real world systems are nonlinear in nature and wide applications are there for nonlinear system identification/modeling.The basic approach in analyzing the nonlinear systems is to build a model from known behavior manifest in the form of system output.The problem of modeling boils down to computing a suitably parameterized model,representing the process.The parameters of the model are adjusted to optimize a performanace function,based on error between the given process output and identified process/model output.While the linear system identification is well established with many classical approaches,most of those methods cannot be directly applied for nonlinear system identification.The problem becomes more complex if the system is completely unknown but only the output time series is available.Blind recognition problem is the direct consequence of such a situation.The thesis concentrates on such problems.Capability of Artificial Neural Networks to approximate many nonlinear input-output maps makes it predominantly suitable for building a function for the identification of nonlinear systems,where only the time series is available.The literature is rich with a variety of algorithms to train the Neural Network model.A comprehensive study of the computation of the model parameters,using the different algorithms and the comparison among them to choose the best technique is still a demanding requirement from practical system designers,which is not available in a concise form in the literature.The thesis is thus an attempt to develop and evaluate some of the well known algorithms and propose some new techniques,in the context of Blind recognition of nonlinear systems.It also attempts to establish the relative merits and demerits of the different approaches.comprehensiveness is achieved in utilizing the benefits of well known evaluation techniques from statistics. The study concludes by providing the results of implementation of the currently available and modified versions and newly introduced techniques for nonlinear blind system modeling followed by a comparison of their performance.It is expected that,such comprehensive study and the comparison process can be of great relevance in many fields including chemical,electrical,biological,financial and weather data analysis.Further the results reported would be of immense help for practical system designers and analysts in selecting the most appropriate method based on the goodness of the model for the particular context.

Investigations On The Development Of An Ann Model & Visual Manipulation Approach For 2-D Dft Computation In Image Processing

This thesis is an outcome of the investigations carried out on the development of an Artificial Neural Network (ANN) model to implement 2-D DFT at high speed. A new definition of 2-D DFT relation is presented. This new definition enables DFT computation organized in stages involving only real addition except at the final stage of computation. The number of stages is always fixed at 4. Two different strategies are proposed. 1) A visual representation of 2-D DFT coefficients. 2) A neural network approach. The visual representation scheme can be used to compute, analyze and manipulate 2D signals such as images in the frequency domain in terms of symbols derived from 2x2 DFT. This, in turn, can be represented in terms of real data. This approach can help analyze signals in the frequency domain even without computing the DFT coefficients. A hierarchical neural network model is developed to implement 2-D DFT. Presently, this model is capable of implementing 2-D DFT for a particular order N such that ((N))4 = 2. The model can be developed into one that can implement the 2-D DFT for any order N upto a set maximum limited by the hardware constraints. The reported method shows a potential in implementing the 2-D DF T in hardware as a VLSI / ASIC

Development of a Biometric Personal Authentication System Based on Fingerprint and Speech

Biometrics deals with the physiological and behavioral characteristics of an individual to establish identity. Fingerprint based authentication is the most advanced biometric authentication technology. The minutiae based fingerprint identification method offer reasonable identification rate. The feature minutiae map consists of about 70-100 minutia points and matching accuracy is dropping down while the size of database is growing up. Hence it is inevitable to make the size of the fingerprint feature code to be as smaller as possible so that identification may be much easier. In this research, a novel global singularity based fingerprint representation is proposed. Fingerprint baseline, which is the line between distal and intermediate phalangeal joint line in the fingerprint, is taken as the reference line. A polygon is formed with the singularities and the fingerprint baseline. The feature vectors are the polygonal angle, sides, area, type and the ridge counts in between the singularities. 100% recognition rate is achieved in this method. The method is compared with the conventional minutiae based recognition method in terms of computation time, receiver operator characteristics (ROC) and the feature vector length. Speech is a behavioural biometric modality and can be used for identification of a speaker. In this work, MFCC of text dependant speeches are computed and clustered using k-means algorithm. A backpropagation based Artificial Neural Network is trained to identify the clustered speech code. The performance of the neural network classifier is compared with the VQ based Euclidean minimum classifier. Biometric systems that use a single modality are usually affected by problems like noisy sensor data, non-universality and/or lack of distinctiveness of the biometric trait, unacceptable error rates, and spoof attacks. Multifinger feature level fusion based fingerprint recognition is developed and the performances are measured in terms of the ROC curve. Score level fusion of fingerprint and speech based recognition system is done and 100% accuracy is achieved for a considerable range of matching threshold

Attribute reduction and missing value imputing with ANN: prediction of learning disabilities

Learning disability (LD) is a neurological condition that affects a child’s brain and impairs his ability to carry out one or many specific tasks. LD affects about 10% of children enrolled in schools. There is no cure for learning disabilities and they are lifelong. The problems of children with specific learning disabilities have been a cause of concern to parents and teachers for some time. Just as there are many different types of LDs, there are a variety of tests that may be done to pinpoint the problem The information gained from an evaluation is crucial for finding out how the parents and the school authorities can provide the best possible learning environment for child. This paper proposes a new approach in artificial neural network (ANN) for identifying LD in children at early stages so as to solve the problems faced by them and to get the benefits to the students, their parents and school authorities. In this study, we propose a closest fit algorithm data preprocessing with ANN classification to handle missing attribute values. This algorithm imputes the missing values in the preprocessing stage. Ignoring of missing attribute values is a common trend in all classifying algorithms. But, in this paper, we use an algorithm in a systematic approach for classification, which gives a satisfactory result in the prediction of LD. It acts as a tool for predicting the LD accurately, and good information of the child is made available to the concerned

Learning disability prediction tool using ANN and ANFIS

Learning Disability (LD) is a neurological condition that affects a child’s brain and impairs his ability to carry out one or many specific tasks. LD affects about 15 % of children enrolled in schools. The prediction of LD is a vital and intricate job. The aim of this paper is to design an effective and powerful tool, using the two intelligent methods viz., Artificial Neural Network and Adaptive Neuro-Fuzzy Inference System, for measuring the percentage of LD that affected in school-age children. In this study, we are proposing some soft computing methods in data preprocessing for improving the accuracy of the tool as well as the classifier. The data preprocessing is performed through Principal Component Analysis for attribute reduction and closest fit algorithm is used for imputing missing values. The main idea in developing the LD prediction tool is not only to predict the LD present in children but also to measure its percentage along with its class like low or minor or major. The system is implemented in Mathworks Software MatLab 7.10. The results obtained from this study have illustrated that the designed prediction system or tool is capable of measuring the LD effectively

On-Line Handwritten Character Recognition using Kohonen Networks

This paper presents an efficient Online Handwritten character Recognition System for Malayalam Characters (OHR-M) using Kohonen network. It would help in recognizing Malayalam text entered using pen-like devices. It will be more natural and efficient way for users to enter text using a pen than keyboard and mouse. To identify the difference between similar characters in Malayalam a novel feature extraction method has been adopted-a combination of context bitmap and normalized (x, y) coordinates. The system reported an accuracy of 88.75% which is writer independent with a recognition time of 15-32 milliseconds

Short-Term Load Forecast Of A Low Loadfactor Power System For Optimization Of Merit Order Dispatch Using Adaptive Learning Algorithm

Short term load forecasting is one of the key inputs to optimize the management of power system. Almost 60-65% of revenue expenditure of a distribution company is against power purchase. Cost of power depends on source of power. Hence any optimization strategy involves optimization in scheduling power from various sources. As the scheduling involves many technical and commercial considerations and constraints, the efficiency in scheduling depends on the accuracy of load forecast. Load forecasting is a topic much visited in research world and a number of papers using different techniques are already presented. The accuracy of forecast for the purpose of merit order dispatch decisions depends on the extent of the permissible variation in generation limits. For a system with low load factor, the peak and the off peak trough are prominent and the forecast should be able to identify these points to more accuracy rather than minimizing the error in the energy content. In this paper an attempt is made to apply Artificial Neural Network (ANN) with supervised learning based approach to make short term load forecasting for a power system with comparatively low load factor. Such power systems are usual in tropical areas with concentrated rainy season for a considerable period of the year

A Reinforcement Learning Approach to Economic Dispatch using Neural Networks

This paper presents a Reinforcement Learning (RL) approach to economic dispatch (ED) using Radial Basis Function neural network. We formulate the ED as an N stage decision making problem. We propose a novel architecture to store Qvalues and present a learning algorithm to learn the weights of the neural network. Even though many stochastic search techniques like simulated annealing, genetic algorithm and evolutionary programming have been applied to ED, they require searching for the optimal solution for each load demand. Also they find limitation in handling stochastic cost functions. In our approach once we learn the Q-values, we can find the dispatch for any load demand. We have recently proposed a RL approach to ED. In that approach, we could find only the optimum dispatch for a set of specified discrete values of power demand. The performance of the proposed algorithm is validated by taking IEEE 6 bus system, considering transmission losses

Soft Computing Approach for Optimization of siRNA Efficiency Prediction for Post-transcriptional Gene Silencing

Post-transcriptional gene silencing by RNA interference is mediated by small interfering RNA called siRNA. This gene silencing mechanism can be exploited therapeutically to a wide variety of disease-associated targets, especially in AIDS, neurodegenerative diseases, cholesterol and cancer on mice with the hope of extending these approaches to treat humans. Over the recent past, a significant amount of work has been undertaken to understand the gene silencing mediated by exogenous siRNA. The design of efficient exogenous siRNA sequences is challenging because of many issues related to siRNA. While designing efficient siRNA, target mRNAs must be selected such that their corresponding siRNAs are likely to be efficient against that target and unlikely to accidentally silence other transcripts due to sequence similarity. So before doing gene silencing by siRNAs, it is essential to analyze their off-target effects in addition to their inhibition efficiency against a particular target. Hence designing exogenous siRNA with good knock-down efficiency and target specificity is an area of concern to be addressed. Some methods have been developed already by considering both inhibition efficiency and off-target possibility of siRNA against agene. Out of these methods, only a few have achieved good inhibition efficiency, specificity and sensitivity. The main focus of this thesis is to develop computational methods to optimize the efficiency of siRNA in terms of “inhibition capacity and off-target possibility” against target mRNAs with improved efficacy, which may be useful in the area of gene silencing and drug design for tumor development. This study aims to investigate the currently available siRNA prediction approaches and to devise a better computational approach to tackle the problem of siRNA efficacy by inhibition capacity and off-target possibility. The strength and limitations of the available approaches are investigated and taken into consideration for making improved solution. Thus the approaches proposed in this study extend some of the good scoring previous state of the art techniques by incorporating machine learning and statistical approaches and thermodynamic features like whole stacking energy to improve the prediction accuracy, inhibition efficiency, sensitivity and specificity. Here, we propose one Support Vector Machine (SVM) model, and two Artificial Neural Network (ANN) models for siRNA efficiency prediction. In SVM model, the classification property is used to classify whether the siRNA is efficient or inefficient in silencing a target gene. The first ANNmodel, named siRNA Designer, is used for optimizing the inhibition efficiency of siRNA against target genes. The second ANN model, named Optimized siRNA Designer, OpsiD, produces efficient siRNAs with high inhibition efficiency to degrade target genes with improved sensitivity-specificity, and identifies the off-target knockdown possibility of siRNA against non-target genes. The models are trained and tested against a large data set of siRNA sequences. The validations are conducted using Pearson Correlation Coefficient, Mathews Correlation Coefficient, Receiver Operating Characteristic analysis, Accuracy of prediction, Sensitivity and Specificity. It is found that the approach, OpsiD, is capable of predicting the inhibition capacity of siRNA against a target mRNA with improved results over the state of the art techniques. Also we are able to understand the influence of whole stacking energy on efficiency of siRNA. The model is further improved by including the ability to identify the “off-target possibility” of predicted siRNA on non-target genes. Thus the proposed model, OpsiD, can predict optimized siRNA by considering both “inhibition efficiency on target genes and off-target possibility on non-target genes”, with improved inhibition efficiency, specificity and sensitivity. Since we have taken efforts to optimize the siRNA efficacy in terms of “inhibition efficiency and offtarget possibility”, we hope that the risk of “off-target effect” while doing gene silencing in various bioinformatics fields can be overcome to a great extent. These findings may provide new insights into cancer diagnosis, prognosis and therapy by gene silencing. The approach may be found useful for designing exogenous siRNA for therapeutic applications and gene silencing techniques in different areas of bioinformatics.

Evolving Distributed Algorithms with Genetic Programming

Distributed systems are one of the most vital components of the economy. The most prominent example is probably the internet, a constituent element of our knowledge society. During the recent years, the number of novel network types has steadily increased. Amongst others, sensor networks, distributed systems composed of tiny computational devices with scarce resources, have emerged. The further development and heterogeneous connection of such systems imposes new requirements on the software development process. Mobile and wireless networks, for instance, have to organize themselves autonomously and must be able to react to changes in the environment and to failing nodes alike. Researching new approaches for the design of distributed algorithms may lead to methods with which these requirements can be met efficiently. In this thesis, one such method is developed, tested, and discussed in respect of its practical utility. Our new design approach for distributed algorithms is based on Genetic Programming, a member of the family of evolutionary algorithms. Evolutionary algorithms are metaheuristic optimization methods which copy principles from natural evolution. They use a population of solution candidates which they try to refine step by step in order to attain optimal values for predefined objective functions. The synthesis of an algorithm with our approach starts with an analysis step in which the wanted global behavior of the distributed system is specified. From this specification, objective functions are derived which steer a Genetic Programming process where the solution candidates are distributed programs. The objective functions rate how close these programs approximate the goal behavior in multiple randomized network simulations. The evolutionary process step by step selects the most promising solution candidates and modifies and combines them with mutation and crossover operators. This way, a description of the global behavior of a distributed system is translated automatically to programs which, if executed locally on the nodes of the system, exhibit this behavior. In our work, we test six different ways for representing distributed programs, comprising adaptations and extensions of well-known Genetic Programming methods (SGP, eSGP, and LGP), one bio-inspired approach (Fraglets), and two new program representations called Rule-based Genetic Programming (RBGP, eRBGP) designed by us. We breed programs in these representations for three well-known example problems in distributed systems: election algorithms, the distributed mutual exclusion at a critical section, and the distributed computation of the greatest common divisor of a set of numbers. Synthesizing distributed programs the evolutionary way does not necessarily lead to the envisaged results. In a detailed analysis, we discuss the problematic features which make this form of Genetic Programming particularly hard. The two Rule-based Genetic Programming approaches have been developed especially in order to mitigate these difficulties. In our experiments, at least one of them (eRBGP) turned out to be a very efficient approach and in most cases, was superior to the other representations.

Das Einzugsgebiet der Fulda - Untersuchung der Methoden der Hochwasserstatistik im Zeichen des Klimawandels

Numerous studies have proven an effect of a probable climate change on the hydrosphere’s different subsystems. In the 21st century global and regional redistribution of water has to be expected and it is very likely that extreme weather phenomenon will occur more frequently. From a global view the flood situation will exacerbate. In contrast to these discoveries the classical approach of flood frequency analysis provides terms like “mean flood recurrence interval”. But for this analysis to be valid there is a need for the precondition of stationary distribution parameters which implies that the flood frequencies are constant in time. Newer approaches take into account extreme value distributions with time-dependent parameters. But the latter implies a discard of the mentioned old terminology that has been used up-to-date in engineering hydrology. On the regional scale climate change affects the hydrosphere in various ways. So, the question appears to be whether in central Europe the classical approach of flood frequency analysis is not usable anymore and whether the traditional terminology should be renewed. In the present case study hydro-meteorological time series of the Fulda catchment area (6930 km²), upstream of the gauging station Bonaforth, are analyzed for the time period 1960 to 2100. At first a distributed catchment area model (SWAT2005) is build up, calibrated and finally validated. The Edertal reservoir is regulated as well by a feedback control of the catchments output in case of low water. Due to this intricacy a special modeling strategy has been necessary: The study area is divided into three SWAT basin models and an additional physically-based reservoir model is developed. To further improve the streamflow predictions of the SWAT model, a correction by an artificial neural network (ANN) has been tested successfully which opens a new way to improve hydrological models. With this extension the calibration and validation of the SWAT model for the Fulda catchment area is improved significantly. After calibration of the model for the past 20th century observed streamflow, the SWAT model is driven by high resolution climate data of the regional model REMO using the IPCC scenarios A1B, A2, and B1, to generate future runoff time series for the 21th century for the various sub-basins in the study area. In a second step flood time series HQ(a) are derived from the 21st century runoff time series (scenarios A1B, A2, and B1). Then these flood projections are extensively tested with regard to stationarity, homogeneity and statistical independence. All these tests indicate that the SWAT-predicted 21st-century trends in the flood regime are not significant. Within the projected time the members of the flood time series are proven to be stationary and independent events. Hence, the classical stationary approach of flood frequency analysis can still be used within the Fulda catchment area, notwithstanding the fact that some regional climate change has been predicted using the IPCC scenarios. It should be noted, however, that the present results are not transferable to other catchment areas. Finally a new method is presented that enables the calculation of extreme flood statistics, even if the flood time series is non-stationary and also if the latter exhibits short- and longterm persistence. This method, which is called Flood Series Maximum Analysis here, enables the calculation of maximum design floods for a given risk- or safety level and time period.