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.

People Recognition in Image Sequences by Supervised Learning

We describe a system that learns from examples to recognize people in images taken indoors. Images of people are represented by color-based and shape-based features. Recognition is carried out through combinations of Support Vector Machine classifiers (SVMs). Different types of multiclass strategies based on SVMs are explored and compared to k-Nearest Neighbors classifiers (kNNs). The system works in real time and shows high performance rates for people recognition throughout one day.

Cross-situational and supervised learning in the emergence of communication

Scenarios for the emergence or bootstrap of a lexicon involve the repeated interaction between at least two agents who must reach a consensus on how to name N objects using H words. Here we consider minimal models of two types of learning algorithms: cross-situational learning, in which the individuals determine the meaning of a word by looking for something in common across all observed uses of that word, and supervised operant conditioning learning, in which there is strong feedback between individuals about the intended meaning of the words. Despite the stark differences between these learning schemes, we show that they yield the same communication accuracy in the limits of large N and H, which coincides with the result of the classical occupancy problem of randomly assigning N objects to H words.

Algoritmos de aprendizado semi-supervisionado baseados em grafos aplicados na bioinformática

Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP)

Semi-supervised dimensionality reduction based on partial least squares for visual analysis of high dimensional data

Dimensionality reduction is employed for visual data analysis as a way to obtaining reduced spaces for high dimensional data or to mapping data directly into 2D or 3D spaces. Although techniques have evolved to improve data segregation on reduced or visual spaces, they have limited capabilities for adjusting the results according to user's knowledge. In this paper, we propose a novel approach to handling both dimensionality reduction and visualization of high dimensional data, taking into account user's input. It employs Partial Least Squares (PLS), a statistical tool to perform retrieval of latent spaces focusing on the discriminability of the data. The method employs a training set for building a highly precise model that can then be applied to a much larger data set very effectively. The reduced data set can be exhibited using various existing visualization techniques. The training data is important to code user's knowledge into the loop. However, this work also devises a strategy for calculating PLS reduced spaces when no training data is available. The approach produces increasingly precise visual mappings as the user feeds back his or her knowledge and is capable of working with small and unbalanced training sets.

Facial Nerve Image Enhancement from CBCT using Supervised Learning Technique

Facial nerve segmentation plays an important role in surgical planning of cochlear implantation. Clinically available CBCT images are used for surgical planning. However, its relatively low resolution renders the identification of the facial nerve difficult. In this work, we present a supervised learning approach to enhance facial nerve image information from CBCT. A supervised learning approach based on multi-output random forest was employed to learn the mapping between CBCT and micro-CT images. Evaluation was performed qualitatively and quantitatively by using the predicted image as input for a previously published dedicated facial nerve segmentation, and cochlear implantation surgical planning software, OtoPlan. Results show the potential of the proposed approach to improve facial nerve image quality as imaged by CBCT and to leverage its segmentation using OtoPlan.

Prediction of Trabecular Bone Anisotropy from Quantitative Computed Tomography using Supervised Learning and a Novel Morphometric Feature Descriptor

Patient-specific biomechanical models including local bone mineral density and anisotropy have gained importance for assessing musculoskeletal disorders. However the trabecular bone anisotropy captured by high-resolution imaging is only available at the peripheral skeleton in clinical practice. In this work, we propose a supervised learning approach to predict trabecular bone anisotropy that builds on a novel set of pose invariant feature descriptors. The statistical relationship between trabecular bone anisotropy and feature descriptors were learned from a database of pairs of high resolution QCT and clinical QCT reconstructions. On a set of leave-one-out experiments, we compared the accuracy of the proposed approach to previous ones, and report a mean prediction error of 6% for the tensor norm, 6% for the degree of anisotropy and 19◦ for the principal tensor direction. These findings show the potential of the proposed approach to predict trabecular bone anisotropy from clinically available QCT images.

Classifying GABAergic interneurons with semi-supervised projected model-based clustering

Objectives: A recently introduced pragmatic scheme promises to be a useful catalog of interneuron names.We sought to automatically classify digitally reconstructed interneuronal morphologies according tothis scheme. Simultaneously, we sought to discover possible subtypes of these types that might emergeduring automatic classification (clustering). We also investigated which morphometric properties weremost relevant for this classification.Materials and methods: A set of 118 digitally reconstructed interneuronal morphologies classified into thecommon basket (CB), horse-tail (HT), large basket (LB), and Martinotti (MA) interneuron types by 42 of theworld?s leading neuroscientists, quantified by five simple morphometric properties of the axon and fourof the dendrites. We labeled each neuron with the type most commonly assigned to it by the experts. Wethen removed this class information for each type separately, and applied semi-supervised clustering tothose cells (keeping the others? cluster membership fixed), to assess separation from other types and lookfor the formation of new groups (subtypes). We performed this same experiment unlabeling the cells oftwo types at a time, and of half the cells of a single type at a time. The clustering model is a finite mixtureof Gaussians which we adapted for the estimation of local (per-cluster) feature relevance. We performedthe described experiments on three different subsets of the data, formed according to how many expertsagreed on type membership: at least 18 experts (the full data set), at least 21 (73 neurons), and at least26 (47 neurons).Results: Interneurons with more reliable type labels were classified more accurately. We classified HTcells with 100% accuracy, MA cells with 73% accuracy, and CB and LB cells with 56% and 58% accuracy,respectively. We identified three subtypes of the MA type, one subtype of CB and LB types each, andno subtypes of HT (it was a single, homogeneous type). We got maximum (adapted) Silhouette widthand ARI values of 1, 0.83, 0.79, and 0.42, when unlabeling the HT, CB, LB, and MA types, respectively,confirming the quality of the formed cluster solutions. The subtypes identified when unlabeling a singletype also emerged when unlabeling two types at a time, confirming their validity. Axonal morphometricproperties were more relevant that dendritic ones, with the axonal polar histogram length in the [pi, 2pi) angle interval being particularly useful.Conclusions: The applied semi-supervised clustering method can accurately discriminate among CB, HT, LB, and MA interneuron types while discovering potential subtypes, and is therefore useful for neuronal classification. The discovery of potential subtypes suggests that some of these types are more heteroge-neous that previously thought. Finally, axonal variables seem to be more relevant than dendritic ones fordistinguishing among the CB, HT, LB, and MA interneuron types.

Semi-supervised construction of general visualization hierarchies

We have recently developed a principled approach to interactive non-linear hierarchical visualization [8] based on the Generative Topographic Mapping (GTM). Hierarchical plots are needed when a single visualization plot is not sufficient (e.g. when dealing with large quantities of data). In this paper we extend our system by giving the user a choice of initializing the child plots of the current plot in either interactive, or automatic mode. In the interactive mode the user interactively selects ``regions of interest'' as in [8], whereas in the automatic mode an unsupervised minimum message length (MML)-driven construction of a mixture of GTMs is used. The latter is particularly useful when the plots are covered with dense clusters of highly overlapping data projections, making it difficult to use the interactive mode. Such a situation often arises when visualizing large data sets. We illustrate our approach on a data set of 2300 18-dimensional points and mention extension of our system to accommodate discrete data types.

Protein-protein interactions classification from text via local learning with class priors

Text classification is essential for narrowing down the number of documents relevant to a particular topic for further pursual, especially when searching through large biomedical databases. Protein-protein interactions are an example of such a topic with databases being devoted specifically to them. This paper proposed a semi-supervised learning algorithm via local learning with class priors (LL-CP) for biomedical text classification where unlabeled data points are classified in a vector space based on their proximity to labeled nodes. The algorithm has been evaluated on a corpus of biomedical documents to identify abstracts containing information about protein-protein interactions with promising results. Experimental results show that LL-CP outperforms the traditional semisupervised learning algorithms such as SVMand it also performs better than local learning without incorporating class priors.

Supervised learning of graph structure

Graph-based representations have been used with considerable success in computer vision in the abstraction and recognition of object shape and scene structure. Despite this, the methodology available for learning structural representations from sets of training examples is relatively limited. In this paper we take a simple yet effective Bayesian approach to attributed graph learning. We present a naïve node-observation model, where we make the important assumption that the observation of each node and each edge is independent of the others, then we propose an EM-like approach to learn a mixture of these models and a Minimum Message Length criterion for components selection. Moreover, in order to avoid the bias that could arise with a single estimation of the node correspondences, we decide to estimate the sampling probability over all the possible matches. Finally we show the utility of the proposed approach on popular computer vision tasks such as 2D and 3D shape recognition. © 2011 Springer-Verlag.

A Multi-View Learning Algorithm

Report published in the Proceedings of the National Conference on "Education and Research in the Information Society", Plovdiv, May, 2014

Application of active and transfer learning to wound image segmentation

Wound management is a fundamental task in standard clinical practice. Automated solutions already exist for humans, but there is a lack of applications on wound management for pets. The importance of a precise and efficient wound assessment is helpful to improve diagnosis and to increase the effectiveness of treatment plans for the chronic wounds. The goal of the research was to propose an automated pipeline capable of segmenting natural light-reflected wound images of animals. Two datasets composed by light-reflected images were used in this work: Deepskin dataset, 1564 human wound images obtained during routine dermatological exams, with 145 manual annotated images; Petwound dataset, a set of 290 wound photos of dogs and cats with 0 annotated images. Two implementations of U-Net Convolutioal Neural Network model were proposed for the automated segmentation. Active Semi-Supervised Learning techniques were applied for human-wound images to perform segmentation from 10% of annotated images. Then the same models were trained, via Transfer Learning, adopting an Active Semi- upervised Learning to unlabelled animal-wound images. The combination of the two training strategies proved their effectiveness in generating large amounts of annotated samples (94% of Deepskin, 80% of PetWound) with the minimal human intervention. The correctness of automated segmentation were evaluated by clinical experts at each round of training thus we can assert that the results obtained in this thesis stands as a reliable solution to perform a correct wound image segmentation. The use of Transfer Learning and Active Semi-Supervied Learning allows to minimize labelling effort from clinicians, even requiring no starting manual annotation at all. Moreover the performances of the model with limited number of parameters suggest the implementation of smartphone-based application to this topic, helping the future standardization of light-reflected images as acknowledge medical images.

Enhancing text classification to improve information filtering

Text classification, information filtering, semi-supervised learning, quality control

Apprentissage machine efficace : théorie et pratique

Malgré des progrès constants en termes de capacité de calcul, mémoire et quantité de données disponibles, les algorithmes d'apprentissage machine doivent se montrer efficaces dans l'utilisation de ces ressources. La minimisation des coûts est évidemment un facteur important, mais une autre motivation est la recherche de mécanismes d'apprentissage capables de reproduire le comportement d'êtres intelligents. Cette thèse aborde le problème de l'efficacité à travers plusieurs articles traitant d'algorithmes d'apprentissage variés : ce problème est vu non seulement du point de vue de l'efficacité computationnelle (temps de calcul et mémoire utilisés), mais aussi de celui de l'efficacité statistique (nombre d'exemples requis pour accomplir une tâche donnée). Une première contribution apportée par cette thèse est la mise en lumière d'inefficacités statistiques dans des algorithmes existants. Nous montrons ainsi que les arbres de décision généralisent mal pour certains types de tâches (chapitre 3), de même que les algorithmes classiques d'apprentissage semi-supervisé à base de graphe (chapitre 5), chacun étant affecté par une forme particulière de la malédiction de la dimensionalité. Pour une certaine classe de réseaux de neurones, appelés réseaux sommes-produits, nous montrons qu'il peut être exponentiellement moins efficace de représenter certaines fonctions par des réseaux à une seule couche cachée, comparé à des réseaux profonds (chapitre 4). Nos analyses permettent de mieux comprendre certains problèmes intrinsèques liés à ces algorithmes, et d'orienter la recherche dans des directions qui pourraient permettre de les résoudre. Nous identifions également des inefficacités computationnelles dans les algorithmes d'apprentissage semi-supervisé à base de graphe (chapitre 5), et dans l'apprentissage de mélanges de Gaussiennes en présence de valeurs manquantes (chapitre 6). Dans les deux cas, nous proposons de nouveaux algorithmes capables de traiter des ensembles de données significativement plus grands. Les deux derniers chapitres traitent de l'efficacité computationnelle sous un angle différent. Dans le chapitre 7, nous analysons de manière théorique un algorithme existant pour l'apprentissage efficace dans les machines de Boltzmann restreintes (la divergence contrastive), afin de mieux comprendre les raisons qui expliquent le succès de cet algorithme. Finalement, dans le chapitre 8 nous présentons une application de l'apprentissage machine dans le domaine des jeux vidéo, pour laquelle le problème de l'efficacité computationnelle est relié à des considérations d'ingénierie logicielle et matérielle, souvent ignorées en recherche mais ô combien importantes en pratique.