Completely supervised training algorithms for B-spline neural networks and neuro-fuzzy systems

Complete supervised training algorithms for B-spline neural networks and fuzzy rule-based systems are discussed. By interducing the relationship between B-spline neural networks and certain types of fuzzy models, training algorithms developed initially for neural networks can be adapted by fuzzy systems.

Supervised training approach using spiking neural networks

One of the basic aspects of some neural networks is their attempt to approximate as much as possible their biological counterparts. The goal is to achieve a simple and robust network, easy to understand and able of simulating the human brain at a computational level. Recently a third generation of neural networks (NN) [1], called Spiking Neural Networks(SNN) was appeared. This new kind of networks use the time of a electrical pulse, or spike, to encode the information. In the first and second generation of NN analog values are used in the communication between neurons.

Unsupervised training for mandarin broadcast news and conversation transcription

A significant cost in obtaining acoustic training data is the generation of accurate transcriptions. For some sources close-caption data is available. This allows the use of lightly-supervised training techniques. However, for some sources and languages close-caption is not available. In these cases unsupervised training techniques must be used. This paper examines the use of unsupervised techniques for discriminative training. In unsupervised training automatic transcriptions from a recognition system are used for training. As these transcriptions may be errorful data selection may be useful. Two forms of selection are described, one to remove non-target language shows, the other to remove segments with low confidence. Experiments were carried out on a Mandarin transcriptions task. Two types of test data were considered, Broadcast News (BN) and Broadcast Conversations (BC). Results show that the gains from unsupervised discriminative training are highly dependent on the accuracy of the automatic transcriptions. © 2007 IEEE.

Combined exercise training in asymptomatic elderly with controlled hypertension: Effects on functional capacity and cardiac diastolic function

Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq)

Supervised stationary cycling versus supervised treadmill walking for patients with intermittent claudication: Preliminary results

Introduction: Walking programmes are recommended as part of the initial treatment for intermittent claudication (IC). However, for many patients factors such as frailty, the severe leg discomfort associated with walking and safety concerns about exercising in public areas reduce compliance to such prescription. Thus, there is a need to identify a mode of exercise that provides the same benefits as regular walking while also offering convenience and comfort for these patients. The present study aims to provide evidence for the first time of the efficacy of a supervised cycle training programme compared with a conventional walking programme for the treatment of IC. Methods: Thus far 33 patients have been randomized to: a treadmill-training group (n = 12); a cycle-training group (n = 11); or a control group (n = 10). Training groups participated in three sessions of supervised training per week for a period of 6 weeks. Control patients received no experimental intervention. Maximal incremental treadmill testing was performed at baseline and after the 6 weeks of training. Measures included pain-free (PFWT) and maximal walking time (MWT), continuous heart rate and gas-analysis recording, and ankle-brachial index assessment. Results: In the treadmill trained group MWT increased significantly from 1016.7 523.7 to 1255.2 432.2 s (P < 0.05). MWT tended to increase with cycle training (848.72 333.18 to 939.54 350.35 s, P = 0.14), and remained unchanged in the control group (1555.1 683.23 to 1534.7 689.87 s). For PFWT, there was a non-significant increase in the treadmill-training group from 414.4 262.3 to 592.9 381.9 s, while both the cycle training and control groups displayed no significant change in this time (226.7 147.1 s to 192.3 56.8 and 499.4 503.7 s to 466.0 526.1 s, respectively). Conclusions: These preliminary results might suggest that, unlike treadmill walking, cycling has no clear effect on walking performance in patients with IC. Thus the current recommendations promoting walking based programmes appear appropriate. The present study was funded by the National Heart Foundation of Australia.

Exponentiated gradient algorithms for large-margin structured classification

We consider the problem of structured classification, where the task is to predict a label y from an input x, and y has meaningful internal structure. Our framework includes supervised training of Markov random fields and weighted context-free grammars as special cases. We describe an algorithm that solves the large-margin optimization problem defined in [12], using an exponential-family (Gibbs distribution) representation of structured objects. The algorithm is efficient—even in cases where the number of labels y is exponential in size—provided that certain expectations under Gibbs distributions can be calculated efficiently. The method for structured labels relies on a more general result, specifically the application of exponentiated gradient updates [7, 8] to quadratic programs.

The effect of exercise intensity on fat oxidation and non-exercise activity thermogenesis in overweight/obese men

It is frequently reported that the actual weight loss achieved through exercise interventions is less than theoretically expected. Amongst other compensatory adjustments that accompany exercise training (e.g., increases in resting metabolic rate and energy intake), a possible cause of the less than expected weight loss is a failure to produce a marked increase in total daily energy expenditure due to a compensatory reduction in non-exercise activity thermogenesis (NEAT). Therefore, there is a need to understand how behaviour is modified in response to exercise interventions. The proposed benefits of exercise training are numerous, including changes to fat oxidation. Given that a diminished capacity to oxidise fat could be a factor in the aetiology of obesity, an exercise training intensity that optimises fat oxidation in overweight/obese individuals would improve impaired fat oxidation, and potentially reduce health risks that are associated with obesity. To improve our understanding of the effectiveness of exercise for weight management, it is important to ensure exercise intensity is appropriately prescribed, and to identify and monitor potential compensatory behavioural changes consequent to exercise training. In line with the gaps in the literature, three studies were performed. The aim of Study 1 was to determine the effect of acute bouts of moderate- and high-intensity walking exercise on NEAT in overweight and obese men. Sixteen participants performed a single bout of either moderate-intensity walking exercise (MIE) or high-intensity walking exercise (HIE) on two separate occasions. The MIE consisted of walking for 60-min on a motorised treadmill at 6 km.h-1. The 60-min HIE session consisted of walking in 5-min intervals at 6 km.h-1 and 10% grade followed by 5-min at 0% grade. NEAT was assessed by accelerometer three days before, on the day of, and three days after the exercise sessions. There was no significant difference in NEAT vector magnitude (counts.min-1) between the pre-exercise period (days 1-3) and the exercise day (day 4) for either protocol. In addition, there was no change in NEAT during the three days following the MIE session, however NEAT increased by 16% on day 7 (post-exercise) compared with the exercise day (P = 0.32). During the post-exercise period following the HIE session, NEAT was increased by 25% on day 7 compared with the exercise day (P = 0.08), and by 30-33% compared with the pre-exercise period (day 1, day 2 and day 3); P = 0.03, 0.03, 0.02, respectively. To conclude, a single bout of either MIE or HIE did not alter NEAT on the exercise day or on the first two days following the exercise session. However, extending the monitoring of NEAT allowed the detection of a 48 hour delay in increased NEAT after performing HIE. A longer-term intervention is needed to determine the effect of accumulated exercise sessions over a week on NEAT. In Study 2, there were two primary aims. The first aim was to test the reliability of a discontinuous incremental exercise protocol (DISCON-FATmax) to identify the workload at which fat oxidation is maximised (FATmax). Ten overweight and obese sedentary male men (mean BMI of 29.5 ¡Ó 4.5 kg/m2 and mean age of 28.0 ¡Ó 5.3 y) participated in this study and performed two identical DISCON-FATmax tests one week apart. Each test consisted of alternate 4-min exercise and 2-min rest intervals on a cycle ergometer. The starting work load of 28 W was increased every 4-min using 14 W increments followed by 2-min rest intervals. When the respiratory exchange ratio was consistently >1.0, the workload was increased by 14 W every 2-min until volitional exhaustion. Fat oxidation was measured by indirect calorimetry. The mean FATmax, ƒtV O2peak, %ƒtV O2peak and %Wmax at which FATmax occurred during the two tests were 0.23 ¡Ó 0.09 and 0.18 ¡Ó 0.08 (g.min-1); 29.7 ¡Ó 7.8 and 28.3 ¡Ó 7.5 (ml.kg-1.min-1); 42.3 ¡Ó 7.2 and 42.6 ¡Ó 10.2 (%ƒtV O2max) and 36.4 ¡Ó 8.5 and 35.4 ¡Ó 10.9 (%), respectively. A paired-samples T-test revealed a significant difference in FATmax (g.min-1) between the tests (t = 2.65, P = 0.03). The mean difference in FATmax was 0.05 (g.min-1) with the 95% confidence interval ranging from 0.01 to 0.18. Paired-samples T-test, however, revealed no significant difference in the workloads (i.e. W) between the tests, t (9) = 0.70, P = 0.4. The intra-class correlation coefficient for FATmax (g.min-1) between the tests was 0.84 (95% confidence interval: 0.36-0.96, P < 0.01). However, Bland-Altman analysis revealed a large disagreement in FATmax (g.min-1) related to W between the two tests; 11 ¡Ó 14 (W) (4.1 ¡Ó 5.3 ƒtV O2peak (%)).These data demonstrate two important phenomena associated with exercise-induced substrate oxidation; firstly, that maximal fat oxidation derived from a discontinuous FATmax protocol differed statistically between repeated tests, and secondly, there was large variability in the workload corresponding with FATmax. The second aim of Study 2 was to test the validity of a DISCON-FATmax protocol by comparing maximal fat oxidation (g.min-1) determined by DISCON-FATmax with fat oxidation (g.min-1) during a continuous exercise protocol using a constant load (CONEX). Ten overweight and obese sedentary males (BMI = 29.5 ¡Ó 4.5 kg/m2; age = 28.0 ¡Ó 4.5 y) with a ƒtV O2max of 29.1 ¡Ó 7.5 ml.kg-1.min-1 performed a DISCON-FATmax test consisting of alternate 4-min exercise and 2-min rest intervals on a cycle ergometer. The 1-h CONEX protocol used the workload from the DISCON-FATmax to determine FATmax. The mean FATmax, ƒtV O2max, %ƒtV O2max and workload at which FATmax occurred during the DISCON-FATmax were 0.23 ¡Ó 0.09 (g.min-1); 29.1 ¡Ó 7.5 (ml.kg-1.min-1); 43.8 ¡Ó 7.3 (%ƒtV O2max) and 58.8 ¡Ó 19.6 (W), respectively. The mean fat oxidation during the 1-h CONEX protocol was 0.19 ¡Ó 0.07 (g.min-1). A paired-samples T-test revealed no significant difference in fat oxidation (g.min-1) between DISCON-FATmax and CONEX, t (9) = 1.85, P = 0.097 (two-tailed). There was also no significant correlation in fat oxidation between the DISCON-FATmax and CONEX (R=0.51, P = 0.14). Bland- Altman analysis revealed a large disagreement in fat oxidation between the DISCONFATmax and CONEX; the upper limit of agreement was 0.13 (g.min-1) and the lower limit of agreement was ¡V0.03 (g.min-1). These data suggest that the CONEX and DISCONFATmax protocols did not elicit different rates of fat oxidation (g.min-1). However, the individual variability in fat oxidation was large, particularly in the DISCON-FATmax test. Further research is needed to ascertain the validity of graded exercise tests for predicting fat oxidation during constant load exercise sessions. The aim of Study 3 was to compare the impact of two different intensities of four weeks of exercise training on fat oxidation, NEAT, and appetite in overweight and obese men. Using a cross-over design 11 participants (BMI = 29 ¡Ó 4 kg/m2; age = 27 ¡Ó 4 y) participated in a training study and were randomly assigned initially to: [1] a lowintensity (45%ƒtV O2max) exercise (LIT) or [2] a high-intensity interval (alternate 30 s at 90%ƒtV O2max followed by 30 s rest) exercise (HIIT) 40-min duration, three times a week. Participants completed four weeks of supervised training and between cross-over had a two week washout period. At baseline and the end of each exercise intervention,ƒtV O2max, fat oxidation, and NEAT were measured. Fat oxidation was determined during a standard 30-min continuous exercise bout at 45%ƒtV O2max. During the steady state exercise expired gases were measured intermittently for 5-min periods and HR was monitored continuously. In each training period, NEAT was measured for seven consecutive days using an accelerometer (RT3) the week before, at week 3 and the week after training. Subjective appetite sensations and food preferences were measured immediately before and after the first exercise session every week for four weeks during both LIT and HIIT. The mean fat oxidation rate during the standard continuous exercise bout at baseline for both LIT and HIIT was 0.14 ¡Ó 0.08 (g.min-1). After four weeks of exercise training, the mean fat oxidation was 0.178 ¡Ó 0.04 and 0.183 ¡Ó 0.04 g.min-1 for LIT and HIIT, respectively. The mean NEAT (counts.min-1) was 45 ¡Ó 18 at baseline, 55 ¡Ó 22 and 44 ¡Ó 16 during training, and 51 ¡Ó 14 and 50 ¡Ó 21 after training for LIT and HIIT, respectively. There was no significant difference in fat oxidation between LIT and HIIT. Moreover, although not statistically significant, there was some evidence to suggest that LIT and HIIT tend to increase fat oxidation during exercise at 45% ƒtV O2max (P = 0.14 and 0.08, respectively). The order of training treatment did not significantly influence changes in fat oxidation, NEAT, and appetite. NEAT (counts.min-1) was not significantly different in the week following training for either LIT or HIIT. Although not statistically significant (P = 0.08), NEAT was 20% lower during week 3 of exercise training in HIIT compared with LIT. Examination of appetite sensations revealed differences in the intensity of hunger, with higher ratings after LIT compared with HIIT. No differences were found in preferences for high-fat sweet foods between LIT and HIIT. In conclusion, the results of this thesis suggest that while fat oxidation during steady state exercise was not affected by the level of exercise intensity, there is strong evidence to suggest that intense exercise could have a debilitative effect on NEAT.

A Prática como Componente Curricular em licenciaturas de Letras: um estudo discursivo das relações entre teoria e prática

O objetivo deste trabalho foi investigar, a partir da reforma das licenciaturas nas universidades ocorrida em 2001, a constituição do eixo disciplinar Prática como Componente Curricular (PCC) nos currículos de licenciatura de cursos de Letras, Português / Espanhol da região sudeste brasileira, focando-nos na relação entre teoria e prática. Para tal discussão, utilizamo-nos das reflexões de Deleuze (1968) a fim de problematizar as possibilidades de repetição total ou de diferenciação total; Vázquez (1977), ao trazer sua discussão de práxis que trata da indissociabilidade da teoria e da prática e Schwartz (2010), para incorporar a discussão sobre o âmbito do trabalho, em particular no que concerne a impossibilidade de antecipação completa da atividade a ser realizada por um profissional. Operamos, também, com os preceitos da Análise do Discurso de base enunciativa (MAINGUENEAU, 1998, 2003) quando tratamos os enunciados como socio-historicamente situados em nossas análises. Para atingirmos nosso fim, realizamos uma contextualização documental que contou com a análise do Parecer CNE/CP 28/2001, no qual estão as determinações sobre carga horária e definição dos eixos de disciplinas da licenciatura, sendo eles: Acadêmico Científico, Prática como Componente Curricular e Estágio Supervisionado. Voltamo-nos, também, para os Projetos Políticos Pedagógicos das universidades analisadas, a fim de investigar qual o entendimento de prática construído nesses documentos. Por fim, recorremos às ementas das disciplinas obrigatórias de PCC oferecidas pelas universidades que compuseram o córpus, buscando identificar as marcas que aproximam a temática da disciplina com o trabalho que considere a prática docente, já que o eixo em questão pressupõe essa discussão. Como critérios de seleção de córpus, consideramos: contemplar uma universidade de cada estado da Região Sudeste; duas universidades que possuem disciplinas que contenham exclusivamente horas de PCC e outras duas que contenham, em uma mesmo disciplina, horas dos eixos de PCC e Acadêmico Científico. Com isso, as universidades analisadas são: UERJ, UFSCar, UFES e UFTM.

Cuidar da pessoa com esquizofrenia em processo de reabilitação psicossocial

Este ensino clínico permitiu-nos prestar cuidados de enfermagem especializados a 16 homens com esquizofrenia em processo de reabilitação psicossocial. Na sua maioria são: adultos, solteiros, da região sul de Portugal, com uma média de 9 anos de escolaridade, com história de consumo de substâncias (álcool, tabaco, haxixe, cocaína, heroína e anfetaminas), institucionalizados na Casa de Saúde do Telhal (CST) há mais de 11 anos. Todos apresentam diagnósticos de enfermagem das várias dimensões do adoecer, com exceção da dimensão comportamental excitatória, sendo os de maior prevalência: “Conhecimento sobre Processo de doença/Cuidados na doença/Tratamento, Não Demonstrado”; Cognição, Comprometida”; “Processo Social, Alterado”; “Autoestima, Diminuída”. Foram desenvolvidos 4 programas de intervenção: consulta de enfermagem de saúde mental e psiquiatria (CESMP); atelier de estimulação cognitiva (AEC); programa de desenvolvimento de competência intrapessoais, interpessoais e profissionais (PDCIIP); programa de psicoeducação (SABER+). A satisfação global com os programas foi superior a 7 (numa escala de 1 a 10) e os resultados que foram percebidos pela maioria dos reabilitandos situaram-se acima das suas expectativas iniciais. De uma forma geral, todos registaram ganhos: no desempenho cognitivo (em particular os que beneficiaram do AEC); na adaptação psicossocial – nomeadamente ao nível do insight; os reabilitandos que beneficiaram do programa SABER+ melhoraram ainda os comportamentos demonstrados de aceitação do estado de saúde; no bem-estar psicológico – nomeadamente ao nível da autoestima (sobretudo os que beneficiaram da CESMP e/ou do PDCIIP), dos afetos positivos e dos comportamentos de motivação; nos vários domínios da qualidade de vida (QdV) medidos pelo WHOQOL-Bref em particular o domínio ambiental; todos elevaram o seu nível de conhecimentos nos 3 eixos em que que se estruturou o programa de psicoeducação (processo da doença, cuidados na doença, tratamento da doença); na adesão à medicação, sobretudo os que integraram o programa de psicoeducação, contudo todos tendem a necessitar de ajuda parcial para conseguir demonstrar conhecimento no cumprimento do esquema terapêutico e estão envolvidos em treinos supervisionados de preparação e autoadministração; ABSTRACT: This clinical training allowed us to provide skilled nursing care to 16 men with schizophrenia in a psychosocial rehabilitation setting. Most of these individuals are: adults, singles, from the southern region of Portugal, with an average of 9 years of schooling, with a history of substance abuse (alcohol, tobacco, cannabis, cocaine, heroin and amphetamines), institutionalized in Casa de Saúde do Telhal (CST) for over 11 years. We find nursing diagnoses from all of the disease dimensions with the exception of excitatory behavioral dimension, being the most prevalent: "knowledge about disease process / care / treatment, not stated"; cognition, impaired", "social process, impaired”, “self-esteem, decreased". We developed 4 intervention programs: psychiatry and mental health nursing consultation (CESMP); atelier of cognitive stimulation (AEC); intrapersonal, interpersonal and professional competences training (PDCIIP); psychoeducation (SABER+). The overall satisfaction with the programs was above 7 (on a scale of 1 to 10) and the results noticed by most patients were above their initial expectations. All registered nursing results were: in cognitive performance (particularly those who benefited from the AEC); in psychosocial adaptation - especially in terms of insight; patients who benefited from the SABER+ program demonstrated improved further acceptance of their health condition, psychological well-being, particularly in terms of self-esteem (especially those who benefited from the CESMP and/or the PDCIIP), positive affects and motivation; in the different domains of quality of life (QoL) measured by WHOQOL-Bref in particular the environmental domain; every patients raised their level of knowledge in the 3 axes in which the psychoeducational program (SABER+) was structured (disease process, care and treatment); in medication adherence, especially those who benefited of the psychoeducational program, however all the patients tend to need partial help to demonstrate knowledge in meeting the therapeutic regimen and are involved in supervised training programs for preparation and self administration.

O discurso do professor no ensino e aprendizagem das equações literais no 8.º ano, no âmbito da experimentação do novo programa de matemática do ensino básico

Relatório da Prática de Ensino Supervisionada, Mestrado em Ensino da Matemática 3.º Ciclo e Secundário, Universidade de Lisboa, 2010

Estimação de cobertura rádio em GSM-R através de redes neuronais

Trabalho Final de Mestrado para obtenção do grau de Mestre em Engenharia de Eletrónica e Telecomunicações

Apprentissage de représentations sur-complètes par entraînement d’auto-encodeurs

Les avancés dans le domaine de l’intelligence artificielle, permettent à des systèmes informatiques de résoudre des tâches de plus en plus complexes liées par exemple à la vision, à la compréhension de signaux sonores ou au traitement de la langue. Parmi les modèles existants, on retrouve les Réseaux de Neurones Artificiels (RNA), dont la popularité a fait un grand bond en avant avec la découverte de Hinton et al. [22], soit l’utilisation de Machines de Boltzmann Restreintes (RBM) pour un pré-entraînement non-supervisé couche après couche, facilitant grandement l’entraînement supervisé du réseau à plusieurs couches cachées (DBN), entraînement qui s’avérait jusqu’alors très difficile à réussir. Depuis cette découverte, des chercheurs ont étudié l’efficacité de nouvelles stratégies de pré-entraînement, telles que l’empilement d’auto-encodeurs traditionnels(SAE) [5, 38], et l’empilement d’auto-encodeur débruiteur (SDAE) [44]. C’est dans ce contexte qu’a débuté la présente étude. Après un bref passage en revue des notions de base du domaine de l’apprentissage machine et des méthodes de pré-entraînement employées jusqu’à présent avec les modules RBM, AE et DAE, nous avons approfondi notre compréhension du pré-entraînement de type SDAE, exploré ses différentes propriétés et étudié des variantes de SDAE comme stratégie d’initialisation d’architecture profonde. Nous avons ainsi pu, entre autres choses, mettre en lumière l’influence du niveau de bruit, du nombre de couches et du nombre d’unités cachées sur l’erreur de généralisation du SDAE. Nous avons constaté une amélioration de la performance sur la tâche supervisée avec l’utilisation des bruits poivre et sel (PS) et gaussien (GS), bruits s’avérant mieux justifiés que celui utilisé jusqu’à présent, soit le masque à zéro (MN). De plus, nous avons démontré que la performance profitait d’une emphase imposée sur la reconstruction des données corrompues durant l’entraînement des différents DAE. Nos travaux ont aussi permis de révéler que le DAE était en mesure d’apprendre, sur des images naturelles, des filtres semblables à ceux retrouvés dans les cellules V1 du cortex visuel, soit des filtres détecteurs de bordures. Nous aurons par ailleurs pu montrer que les représentations apprises du SDAE, composées des caractéristiques ainsi extraites, s’avéraient fort utiles à l’apprentissage d’une machine à vecteurs de support (SVM) linéaire ou à noyau gaussien, améliorant grandement sa performance de généralisation. Aussi, nous aurons observé que similairement au DBN, et contrairement au SAE, le SDAE possédait une bonne capacité en tant que modèle générateur. Nous avons également ouvert la porte à de nouvelles stratégies de pré-entraînement et découvert le potentiel de l’une d’entre elles, soit l’empilement d’auto-encodeurs rebruiteurs (SRAE).

Recognition of Facial Expressions with Autoencoders and Convolutional-Nets

Les humains communiquent via différents types de canaux: les mots, la voix, les gestes du corps, des émotions, etc. Pour cette raison, un ordinateur doit percevoir ces divers canaux de communication pour pouvoir interagir intelligemment avec les humains, par exemple en faisant usage de microphones et de webcams. Dans cette thèse, nous nous intéressons à déterminer les émotions humaines à partir d’images ou de vidéo de visages afin d’ensuite utiliser ces informations dans différents domaines d’applications. Ce mémoire débute par une brève introduction à l'apprentissage machine en s’attardant aux modèles et algorithmes que nous avons utilisés tels que les perceptrons multicouches, réseaux de neurones à convolution et autoencodeurs. Elle présente ensuite les résultats de l'application de ces modèles sur plusieurs ensembles de données d'expressions et émotions faciales. Nous nous concentrons sur l'étude des différents types d’autoencodeurs (autoencodeur débruitant, autoencodeur contractant, etc) afin de révéler certaines de leurs limitations, comme la possibilité d'obtenir de la coadaptation entre les filtres ou encore d’obtenir une courbe spectrale trop lisse, et étudions de nouvelles idées pour répondre à ces problèmes. Nous proposons également une nouvelle approche pour surmonter une limite des autoencodeurs traditionnellement entrainés de façon purement non-supervisée, c'est-à-dire sans utiliser aucune connaissance de la tâche que nous voulons finalement résoudre (comme la prévision des étiquettes de classe) en développant un nouveau critère d'apprentissage semi-supervisé qui exploite un faible nombre de données étiquetées en combinaison avec une grande quantité de données non-étiquetées afin d'apprendre une représentation adaptée à la tâche de classification, et d'obtenir une meilleure performance de classification. Finalement, nous décrivons le fonctionnement général de notre système de détection d'émotions et proposons de nouvelles idées pouvant mener à de futurs travaux.

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.