Redes neurais pulsadas - redes que utilizam uma codificação temporal da informação - têm despontado como uma nova e promissora abordagem dentro do paradigma conexionista emergente da ciência cognitiva. Um desses novos modelos é a rede neural pulsada de base radial, capaz de armazenar informação nos tempos de atraso axonais dos neurônios e que comporta algoritmos explícitos de treinamento. A recente proposição de uma sistemática para a codificação temporal dos dados de entrada utilizando campos receptivos gaussianos tem apresentado interessantes resultados na tarefa do agrupamento de dados (clustering). Este trabalho propõe uma função para o aprendizado não supervisionado dessa rede, com o objetivo de simplificar a sistemática de calibração de alguns dos seus parâmetros-chave, aprimorando a convergência da rede neural pulsada no aprendizado baseado em instâncias. O desempenho desse modelo é avaliado na tarefa de classificação de padrões, particularmente na classificação de pixels em imagens coloridas no domínio da visão computacional. / Pulsed neural networks - networks that encode information in the timing of spikes - have been studied as a new and promising approach in the artificial neural networks paradigm, emergent from cognitive science. One of these new models is the pulsed neural network with radial basis function, a network able to store information in the axonal propagation delay of neurons. Recently, a new method for encoding input-data by population code using gaussian receptive fields has showed interesting results in the clustering task. The present work proposes a function for the unsupervised learning task in this network, which goal includes the simplification of the calibration of the network key parameters and the enhancement of the pulsed neural network convergence to instance based learning. The performance of this model is evaluated for pattern classification, particularly for the pixel colors classification task, in the computer vision domain.

Diese Doktorarbeit führt ein hierarchisches Modell für das unüberwachte Lernen aus quasi-natürlichen Videosequenzen ein. Das Modell basiert auf den Lernprinzipien der Langsamkeit und Spärlichkeit, für die verschiedene Ansätze und Implementierungen vorgestellt werden. Eine Vielzahl von Neuronentypen im Hippocampus von Nagern und Primaten kodiert verschiedene Aspekte der räumlichen Umgebung eines Tieres. Dazu gehören Ortszellen (place cells), Kopfrichtungszellen (head direction cells), Raumansichtszellen (spatial view cells) und Gitterzellen (grid cells). Die Hauptergebnisse dieser Arbeit basieren auf dem Training des hierarchischen Modells mit Videosequenzen aus einer Virtual-Reality-Umgebung. Das Modell reproduziert die wichtigsten räumlichen Codes aus dem Hippocampus. Die Art der erzeugten Repräsentationen hängt hauptsächlich von der Bewegungsstatistik des simulierten Tieres ab. Das vorgestellte Modell wird außerdem auf das Problem der invaranten Objekterkennung angewandt, indem Videosequenzen von simulierten Kugelhaufen oder Fischen als Stimuli genutzt wurden. Die resultierenden Modellrepräsentationen erlauben das unabhängige Auslesen von Objektidentität, Position und Rotationswinkel im Raum. / This thesis introduces a hierarchical model for unsupervised learning from naturalistic video sequences. The model is based on the principles of slowness and sparseness. Different approaches and implementations for these principles are discussed. A variety of neuron classes in the hippocampal formation of rodents and primates codes for different aspects of space surrounding the animal, including place cells, head direction cells, spatial view cells and grid cells. In the main part of this thesis, video sequences from a virtual reality environment are used for training the hierarchical model. The behavior of most known hippocampal neuron types coding for space are reproduced by this model. The type of representations generated by the model is mostly determined by the movement statistics of the simulated animal. The model approach is not limited to spatial coding. An application of the model to invariant object recognition is described, where artificial clusters of spheres or rendered fish are presented to the model. The resulting representations allow a simple extraction of the identity of the object presented as well as of its position and viewing angle.

No abstract

In this thesis we develop a class of nonlinear generative models for high-dimensional time series. The first key property of these models is their distributed, or "componential" latent state, which is characterized by binary stochastic variables which interact to explain the data. The second key property is the use of an undirected graphical model to represent the relationship between latent state (features) and observations. The final key property is composability: the proposed class of models can form the building blocks of deep networks by successively training each model on the features extracted by the previous one. We first propose a model based on the Restricted Boltzmann Machine (RBM) that uses an undirected model with binary latent variables and real-valued "visible" variables. The latent and visible variables at each time step receive directed connections from the visible variables at the last few time-steps. This "conditional" RBM (CRBM) makes on-line inference efficient and allows us to use a simple approximate learning procedure. We demonstrate the power of our approach by synthesizing various motion sequences and by performing on-line filling in of data lost during motion capture. We also explore CRBMs as priors in the context of Bayesian filtering applied to multi-view and monocular 3D person tracking. We extend the CRBM in a way that preserves its most important computational properties and introduces multiplicative three-way interactions that allow the effective interaction weight between two variables to be modulated by the dynamic state of a third variable. We introduce a factoring of the implied three-way weight tensor to permit a more compact parameterization. The resulting model can capture diverse styles of motion with a single set of parameters, and the three-way interactions greatly improve its ability to blend motion styles or to transition smoothly among them. In separate but related work, we revisit Products of Hidden Markov Models (PoHMMs). We show how the partition function can be estimated reliably via Annealed Importance Sampling. This enables us to demonstrate that PoHMMs outperform various flavours of HMMs on a variety of tasks and metrics, including log likelihood.

This dissertation focuses on improving content-based recommendation systems for music. Specifically, progress in the development in music content-based recommendation systems has stalled in recent years due to some faulty assumptions: 1. most acoustic content-based systems for music information retrieval (MIR) assume a bag-of-frames model, where it is assumed that a song contains a simplistic, global audio texture 2. genre, style, mood, and authors are appropriate categories for machine-oriented recommendation 3. similarity is a universal construct and does not vary among different users The main contribution of this dissertation is to address these faulty assumptions by describing a novel approach in MIR that provides user-centric, content-based recommendations based on statistics of acoustic sound elements. First, this dissertation presents the acoustic segment modeling framework that describes a piece of music as a temporal sequence of acoustic segment models (ASMs), which represent individual polyphonic sound elements. A dictionary of ASMs generated in an unsupervised process defines a vocabulary of acoustic tokens that are able to transcribe new musical pieces. Next, standard text-based information retrieval algorithms use statistics of ASM counts to perform various retrieval tasks. Despite a simple feature set compared to other content-based genre recommendation algorithms, the acoustic segment modeling approach is highly competitive on standard genre classification databases. Fundamental to the success of the acoustic segment modeling approach is the ability to model acoustical semantics in a musical piece, which is demonstrated by the detection of musical attributes on temporal characteristics. Further, it is shown that the acoustic segment modeling procedure is able to capture the inherent structure of melody by providing near state-of-the-art performance on an automatic chord recognition task. This dissertation demonstrates that some classification tasks, such as genre, possess information that is not contained in the acoustic signal; therefore, attempts at modeling these categories using only the acoustic content is ill-fated. Further, notions of music similarity are personal in nature and are not derived from a universal ontology. Therefore, this dissertation addresses the second and third limitation of previous content-based retrieval approaches by presenting a user-centric preference rating algorithm. Individual users possess their own cognitive construct of similarity; therefore, retrieval algorithms must demonstrate this flexibility. The proposed rating algorithm is based on the principle of minimum classification error (MCE) training, which has been demonstrated to be robust against outliers and also minimizes the Parzen estimate of the theoretical classification risk. The outlier immunity property limits the effect of labels that arise from non-content-based sources. The MCE-based algorithm performs better than a similar ratings prediction algorithm. Further, this dissertation discusses extensions and future work.

In this thesis we develop a class of nonlinear generative models for high-dimensional time series. The first key property of these models is their distributed, or "componential" latent state, which is characterized by binary stochastic variables which interact to explain the data. The second key property is the use of an undirected graphical model to represent the relationship between latent state (features) and observations. The final key property is composability: the proposed class of models can form the building blocks of deep networks by successively training each model on the features extracted by the previous one. We first propose a model based on the Restricted Boltzmann Machine (RBM) that uses an undirected model with binary latent variables and real-valued "visible" variables. The latent and visible variables at each time step receive directed connections from the visible variables at the last few time-steps. This "conditional" RBM (CRBM) makes on-line inference efficient and allows us to use a simple approximate learning procedure. We demonstrate the power of our approach by synthesizing various motion sequences and by performing on-line filling in of data lost during motion capture. We also explore CRBMs as priors in the context of Bayesian filtering applied to multi-view and monocular 3D person tracking. We extend the CRBM in a way that preserves its most important computational properties and introduces multiplicative three-way interactions that allow the effective interaction weight between two variables to be modulated by the dynamic state of a third variable. We introduce a factoring of the implied three-way weight tensor to permit a more compact parameterization. The resulting model can capture diverse styles of motion with a single set of parameters, and the three-way interactions greatly improve its ability to blend motion styles or to transition smoothly among them. In separate but related work, we revisit Products of Hidden Markov Models (PoHMMs). We show how the partition function can be estimated reliably via Annealed Importance Sampling. This enables us to demonstrate that PoHMMs outperform various flavours of HMMs on a variety of tasks and metrics, including log likelihood.

Thèse numérisée par la Division de la gestion de documents et des archives de l'Université de Montréal

Tandem mass spectrometry (MS/MS) is the dominant approach for large-scale peptide sequencing in high-throughput proteomic profiling studies. The computational analysis of MS/MS spectra involves the identification of peptides from experimental spectra, especially those with post-translational modifications (PTMs), as well as the inference of protein composition based on the putative identified peptides. In this thesis, we tackled two major challenges associated with an MS/MS analysis: 1) the refinement of PTM predictions from MS/MS spectra and 2) the inference of protein composition based on peptide predictions. We proposed two PTM prediction refinement algorithms, PTMClust and its Bayesian nonparametric extension \emph{i}PTMClust, and a protein identification algorithm, pro-HAP, that is based on a novel two-layer hierarchical clustering approach that leverages prior knowledge about protein function. Individually, we show that our two PTM refinement algorithms outperform the state-of-the-art algorithms and our protein identification algorithm performs at par with the state of the art. Collectively, as a demonstration of our end-to-end MS/MS computational analysis of a human chromatin protein complex study, we show that our analysis pipeline can find high confidence putative novel protein complex members. Moreover, it can provide valuable insights into the formation and regulation of protein complexes by detailing the specificity of different PTMs for the members in each complex.

Tandem mass spectrometry (MS/MS) is the dominant approach for large-scale peptide sequencing in high-throughput proteomic profiling studies. The computational analysis of MS/MS spectra involves the identification of peptides from experimental spectra, especially those with post-translational modifications (PTMs), as well as the inference of protein composition based on the putative identified peptides. In this thesis, we tackled two major challenges associated with an MS/MS analysis: 1) the refinement of PTM predictions from MS/MS spectra and 2) the inference of protein composition based on peptide predictions. We proposed two PTM prediction refinement algorithms, PTMClust and its Bayesian nonparametric extension \emph{i}PTMClust, and a protein identification algorithm, pro-HAP, that is based on a novel two-layer hierarchical clustering approach that leverages prior knowledge about protein function. Individually, we show that our two PTM refinement algorithms outperform the state-of-the-art algorithms and our protein identification algorithm performs at par with the state of the art. Collectively, as a demonstration of our end-to-end MS/MS computational analysis of a human chromatin protein complex study, we show that our analysis pipeline can find high confidence putative novel protein complex members. Moreover, it can provide valuable insights into the formation and regulation of protein complexes by detailing the specificity of different PTMs for the members in each complex.

Recent advancements in Artificial Intelligence has prompted a shearexplosion of new research initiatives and applications, improving notonly existing technologies, but also opening up opportunities for newand exiting applications. This thesis explores the MIMIC-III intensive care unit database and conducts experiment on an interpretable feature space based on sever-ty scores, defining a patient health state, commonly used to predict mortality in an ICU setting. Patient health state trajectories are clustered and correlated with administered medication and performed procedures to get a better understanding of the potential usefulness in evaluating treatments on their effect on said health state, where commonalities and deviations in treatment can be understood. Furthermore, medication and procedure classification is carried out to explore their predictability using the severity subscore feature space.