Sparse Representations of Image via Overcomplete Dictionary Learned by Adaptive Non-orthogonal Sparsifying Transform

How to learn an over complete dictionary for sparse representations of image is an important topic in machine learning, sparse coding, blind source separation, etc. The so-called K-singular value decomposition (K-SVD) method [3] is powerful for this purpose, however, it is too time-consuming to apply. Recently, an adaptive orthogonal sparsifying transform (AOST) method has been developed to learn the dictionary that is faster. However, the corresponding coefficient matrix may not be as sparse as that of K-SVD. For solving this problem, in this paper, a non-orthogonal iterative match method is proposed to learn the dictionary. By using the approach of sequentially extracting columns of the stacked image blocks, the non-orthogonal atoms of the dictionary are learned adaptively, and the resultant coefficient matrix is sparser. Experiment results show that the proposed method can yield effective dictionaries and the resulting image representation is sparser than AOST.

Sparse Signal Representation in Digital and Biological Systems

Theories of sparse signal representation, wherein a signal is decomposed as the sum of a small number of constituent elements, play increasing roles in both mathematical signal processing and neuroscience. This happens despite the differences between signal models in the two domains. After reviewing preliminary material on sparse signal models, I use work on compressed sensing for the electron tomography of biological structures as a target for exploring the efficacy of sparse signal reconstruction in a challenging application domain. My research in this area addresses a topic of keen interest to the biological microscopy community, and has resulted in the development of tomographic reconstruction software which is competitive with the state of the art in its field. Moving from the linear signal domain into the nonlinear dynamics of neural encoding, I explain the sparse coding hypothesis in neuroscience and its relationship with olfaction in locusts. I implement a numerical ODE model of the activity of neural populations responsible for sparse odor coding in locusts as part of a project involving offset spiking in the Kenyon cells. I also explain the validation procedures we have devised to help assess the model's similarity to the biology. The thesis concludes with the development of a new, simplified model of locust olfactory network activity, which seeks with some success to explain statistical properties of the sparse coding processes carried out in the network.

Random projections on manifolds of symmetric positive definite matrices for image classification

Recent advances suggest that encoding images through Symmetric Positive Definite (SPD) matrices and then interpreting such matrices as points on Riemannian manifolds can lead to increased classification performance. Taking into account manifold geometry is typically done via (1) embedding the manifolds in tangent spaces, or (2) embedding into Reproducing Kernel Hilbert Spaces (RKHS). While embedding into tangent spaces allows the use of existing Euclidean-based learning algorithms, manifold shape is only approximated which can cause loss of discriminatory information. The RKHS approach retains more of the manifold structure, but may require non-trivial effort to kernelise Euclidean-based learning algorithms. In contrast to the above approaches, in this paper we offer a novel solution that allows SPD matrices to be used with unmodified Euclidean-based learning algorithms, with the true manifold shape well-preserved. Specifically, we propose to project SPD matrices using a set of random projection hyperplanes over RKHS into a random projection space, which leads to representing each matrix as a vector of projection coefficients. Experiments on face recognition, person re-identification and texture classification show that the proposed approach outperforms several recent methods, such as Tensor Sparse Coding, Histogram Plus Epitome, Riemannian Locality Preserving Projection and Relational Divergence Classification.

Human action recognition from video sequences

This PhD research has proposed new machine learning techniques to improve human action recognition based on local features. Several novel video representation and classification techniques have been proposed to increase the performance with lower computational complexity. The major contributions are the construction of new feature representation techniques, based on advanced machine learning techniques such as multiple instance dictionary learning, Latent Dirichlet Allocation (LDA) and Sparse coding. A Binary-tree based classification technique was also proposed to deal with large amounts of action categories. These techniques are not only improving the classification accuracy with constrained computational resources but are also robust to challenging environmental conditions. These developed techniques can be easily extended to a wide range of video applications to provide near real-time performance.

Dictionary-Learning-Based Post-Filter for HMM-Based Speech Synthesis

Oversmoothing of speech parameter trajectories is one of the causes for quality degradation of HMM-based speech synthesis. Various methods have been proposed to overcome this effect, the most recent ones being global variance (GV) and modulation-spectrum-based post-filter (MSPF). However, there is still a significant quality gap between natural and synthesized speech. In this paper, we propose a two-fold post-filtering technique to alleviate to a certain extent the oversmoothing of spectral and excitation parameter trajectories of HMM-based speech synthesis. For the spectral parameters, we propose a sparse coding-based post-filter to match the trajectories of synthetic speech to that of natural speech, and for the excitation trajectory, we introduce a perceptually motivated post-filter. Experimental evaluations show quality improvement compared with existing methods.

On sparsity and overcompleteness in image models

Computational models of visual cortex, and in particular those based on sparse coding, have enjoyed much recent attention. Despite this currency, the question of how sparse or how over-complete a sparse representation should be, has gone without principled answer. Here, we use Bayesian model-selection methods to address these questions for a sparse-coding model based on a Student-t prior. Having validated our methods on toy data, we find that natural images are indeed best modelled by extremely sparse distributions; although for the Student-t prior, the associated optimal basis size is only modestly over-complete.

Slowness and sparseness lead to place, head-direction, and spatial-view cells.

We present a model for the self-organized formation of place cells, head-direction cells, and spatial-view cells in the hippocampal formation based on unsupervised learning on quasi-natural visual stimuli. The model comprises a hierarchy of Slow Feature Analysis (SFA) nodes, which were recently shown to reproduce many properties of complex cells in the early visual system []. The system extracts a distributed grid-like representation of position and orientation, which is transcoded into a localized place-field, head-direction, or view representation, by sparse coding. The type of cells that develops depends solely on the relevant input statistics, i.e., the movement pattern of the simulated animal. The numerical simulations are complemented by a mathematical analysis that allows us to accurately predict the output of the top SFA layer.

A novel shrinkage technique based on the normal inverse Gaussian density model

This paper proposes a novel image denoising technique based on the normal inverse Gaussian (NIG) density model using an extended non-negative sparse coding (NNSC) algorithm proposed by us. This algorithm can converge to feature basis vectors, which behave in the locality and orientation in spatial and frequency domain. Here, we demonstrate that the NIG density provides a very good fitness to the non-negative sparse data. In the denoising process, by exploiting a NIG-based maximum a posteriori estimator (MAP) of an image corrupted by additive Gaussian noise, the noise can be reduced successfully. This shrinkage technique, also referred to as the NNSC shrinkage technique, is self-adaptive to the statistical properties of image data. This denoising method is evaluated by values of the normalized signal to noise rate (SNR). Experimental results show that the NNSC shrinkage approach is indeed efficient and effective in denoising. Otherwise, we also compare the effectiveness of the NNSC shrinkage method with methods of standard sparse coding shrinkage, wavelet-based shrinkage and the Wiener filter. The simulation results show that our method outperforms the three kinds of denoising approaches mentioned above.

Non-negative matrix decomposition approaches to frequency domain analysis of music audio signals

On étudie l’application des algorithmes de décomposition matricielles tel que la Factorisation Matricielle Non-négative (FMN), aux représentations fréquentielles de signaux audio musicaux. Ces algorithmes, dirigés par une fonction d’erreur de reconstruction, apprennent un ensemble de fonctions de base et un ensemble de coef- ficients correspondants qui approximent le signal d’entrée. On compare l’utilisation de trois fonctions d’erreur de reconstruction quand la FMN est appliquée à des gammes monophoniques et harmonisées: moindre carré, divergence Kullback-Leibler, et une mesure de divergence dépendente de la phase, introduite récemment. Des nouvelles méthodes pour interpréter les décompositions résultantes sont présentées et sont comparées aux méthodes utilisées précédemment qui nécessitent des connaissances du domaine acoustique. Finalement, on analyse la capacité de généralisation des fonctions de bases apprises par rapport à trois paramètres musicaux: l’amplitude, la durée et le type d’instrument. Pour ce faire, on introduit deux algorithmes d’étiquetage des fonctions de bases qui performent mieux que l’approche précédente dans la majorité de nos tests, la tâche d’instrument avec audio monophonique étant la seule exception importante.

Deep learning of representations and its application to computer vision

L’objectif de cette thèse par articles est de présenter modestement quelques étapes du parcours qui mènera (on espère) à une solution générale du problème de l’intelligence artificielle. Cette thèse contient quatre articles qui présentent chacun une différente nouvelle méthode d’inférence perceptive en utilisant l’apprentissage machine et, plus particulièrement, les réseaux neuronaux profonds. Chacun de ces documents met en évidence l’utilité de sa méthode proposée dans le cadre d’une tâche de vision par ordinateur. Ces méthodes sont applicables dans un contexte plus général, et dans certains cas elles on tété appliquées ailleurs, mais ceci ne sera pas abordé dans le contexte de cette de thèse. Dans le premier article, nous présentons deux nouveaux algorithmes d’inférence variationelle pour le modèle génératif d’images appelé codage parcimonieux “spike- and-slab” (CPSS). Ces méthodes d’inférence plus rapides nous permettent d’utiliser des modèles CPSS de tailles beaucoup plus grandes qu’auparavant. Nous démontrons qu’elles sont meilleures pour extraire des détecteur de caractéristiques quand très peu d’exemples étiquetés sont disponibles pour l’entraînement. Partant d’un modèle CPSS, nous construisons ensuite une architecture profonde, la machine de Boltzmann profonde partiellement dirigée (MBP-PD). Ce modèle a été conçu de manière à simplifier d’entraînement des machines de Boltzmann profondes qui nécessitent normalement une phase de pré-entraînement glouton pour chaque couche. Ce problème est réglé dans une certaine mesure, mais le coût d’inférence dans le nouveau modèle est relativement trop élevé pour permettre de l’utiliser de manière pratique. Dans le deuxième article, nous revenons au problème d’entraînement joint de machines de Boltzmann profondes. Cette fois, au lieu de changer de famille de modèles, nous introduisons un nouveau critère d’entraînement qui donne naissance aux machines de Boltzmann profondes à multiples prédictions (MBP-MP). Les MBP-MP sont entraînables en une seule étape et ont un meilleur taux de succès en classification que les MBP classiques. Elles s’entraînent aussi avec des méthodes variationelles standard au lieu de nécessiter un classificateur discriminant pour obtenir un bon taux de succès en classification. Par contre, un des inconvénients de tels modèles est leur incapacité de générer deséchantillons, mais ceci n’est pas trop grave puisque la performance de classification des machines de Boltzmann profondes n’est plus une priorité étant donné les dernières avancées en apprentissage supervisé. Malgré cela, les MBP-MP demeurent intéressantes parce qu’elles sont capable d’accomplir certaines tâches que des modèles purement supervisés ne peuvent pas faire, telles que celle de classifier des données incomplètes ou encore celle de combler intelligemment l’information manquante dans ces données incomplètes. Le travail présenté dans cette thèse s’est déroulé au milieu d’une période de transformations importantes du domaine de l’apprentissage à réseaux neuronaux profonds qui a été déclenchée par la découverte de l’algorithme de “dropout” par Geoffrey Hinton. Dropout rend possible un entraînement purement supervisé d’architectures de propagation unidirectionnel sans être exposé au danger de sur- entraînement. Le troisième article présenté dans cette thèse introduit une nouvelle fonction d’activation spécialement con ̧cue pour aller avec l’algorithme de Dropout. Cette fonction d’activation, appelée maxout, permet l’utilisation de aggrégation multi-canal dans un contexte d’apprentissage purement supervisé. Nous démontrons comment plusieurs tâches de reconnaissance d’objets sont mieux accomplies par l’utilisation de maxout. Pour terminer, sont présentons un vrai cas d’utilisation dans l’industrie pour la transcription d’adresses de maisons à plusieurs chiffres. En combinant maxout avec une nouvelle sorte de couche de sortie pour des réseaux neuronaux de convolution, nous démontrons qu’il est possible d’atteindre un taux de succès comparable à celui des humains sur un ensemble de données coriace constitué de photos prises par les voitures de Google. Ce système a été déployé avec succès chez Google pour lire environ cent million d’adresses de maisons.

Low rank representation on Riemannian manifold of symmetric positive definite matrices

Sparse coding aims to find a more compact representation based on a set of dictionary atoms. A well-known technique looking at 2D sparsity is the low rank representation (LRR). However, in many computer vision applications, data often originate from a manifold, which is equipped with some Riemannian geometry. In this case, the existing LRR becomes inappropriate for modeling and incorporating the intrinsic geometry of the manifold that is potentially important and critical to applications. In this paper, we generalize the LRR over the Euclidean space to the LRR model over a specific Rimannian manifold—the manifold of symmetric positive matrices (SPD). Experiments on several computer vision datasets showcase its noise robustness and superior performance on classification and segmentation compared with state-of-the-art approaches.

A biologically inspired appearance model for robust visual tracking

In this work, we propose a biologically inspired appearance model for robust visual tracking. Motivated in part by the success of the hierarchical organization of the primary visual cortex (area V1), we establish an architecture consisting of five layers: whitening, rectification, normalization, coding and polling. The first three layers stem from the models developed for object recognition. In this paper, our attention focuses on the coding and pooling layers. In particular, we use a discriminative sparse coding method in the coding layer along with spatial pyramid representation in the pooling layer, which makes it easier to distinguish the target to be tracked from its background in the presence of appearance variations. An extensive experimental study shows that the proposed method has higher tracking accuracy than several state-of-the-art trackers.

Using electroencephalography signals to control acoustical processing

Current hearing-assistive technology performs poorly in noisy multi-talker conditions. The goal of this thesis was to establish the feasibility of using EEG to guide acoustic processing in such conditions. To attain this goal, this research developed a model via the constructive research method, relying on literature review. Several approaches have revealed improvements in the performance of hearing-assistive devices under multi-talker conditions, namely beamforming spatial filtering, model-based sparse coding shrinkage, and onset enhancement of the speech signal. Prior research has shown that electroencephalography (EEG) signals contain information that concerns whether the person is actively listening, what the listener is listening to, and where the attended sound source is. This thesis constructed a model for using EEG information to control beamforming, model-based sparse coding shrinkage, and onset enhancement of the speech signal. The purpose of this model is to propose a framework for using EEG signals to control sound processing to select a single talker in a noisy environment containing multiple talkers speaking simultaneously. On a theoretical level, the model showed that EEG can control acoustical processing. An analysis of the model identified a requirement for real-time processing and that the model inherits the computationally intensive properties of acoustical processing, although the model itself is low complexity placing a relatively small load on computational resources. A research priority is to develop a prototype that controls hearing-assistive devices with EEG. This thesis concludes highlighting challenges for future research.

Generating Discriminative Objective Proposals via Submodular Ranking

Object recognition has long been a core problem in computer vision. To improve object spatial support and speed up object localization for object recognition, generating high-quality category-independent object proposals as the input for object recognition system has drawn attention recently. Given an image, we generate a limited number of high-quality and category-independent object proposals in advance and used as inputs for many computer vision tasks. We present an efficient dictionary-based model for image classification task. We further extend the work to a discriminative dictionary learning method for tensor sparse coding. In the first part, a multi-scale greedy-based object proposal generation approach is presented. Based on the multi-scale nature of objects in images, our approach is built on top of a hierarchical segmentation. We first identify the representative and diverse exemplar clusters within each scale. Object proposals are obtained by selecting a subset from the multi-scale segment pool via maximizing a submodular objective function, which consists of a weighted coverage term, a single-scale diversity term and a multi-scale reward term. The weighted coverage term forces the selected set of object proposals to be representative and compact; the single-scale diversity term encourages choosing segments from different exemplar clusters so that they will cover as many object patterns as possible; the multi-scale reward term encourages the selected proposals to be discriminative and selected from multiple layers generated by the hierarchical image segmentation. The experimental results on the Berkeley Segmentation Dataset and PASCAL VOC2012 segmentation dataset demonstrate the accuracy and efficiency of our object proposal model. Additionally, we validate our object proposals in simultaneous segmentation and detection and outperform the state-of-art performance. To classify the object in the image, we design a discriminative, structural low-rank framework for image classification. We use a supervised learning method to construct a discriminative and reconstructive dictionary. By introducing an ideal regularization term, we perform low-rank matrix recovery for contaminated training data from all categories simultaneously without losing structural information. A discriminative low-rank representation for images with respect to the constructed dictionary is obtained. With semantic structure information and strong identification capability, this representation is good for classification tasks even using a simple linear multi-classifier.