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.

The use of social and salience cues in early word learning

This article explores young infants' ability to learn new words in situations providing tightly controlled social and salience cues to their reference. Four experiments investigated whether, given two potential referents, 15-month-olds would attach novel labels to (a) an image toward which a digital recording of a face turned and gazed, (b) a moving image versus a stationary image, (c) a moving image toward which the face gazed, and (d) a gazed-on image versus a moving image. Infants successfully used the recorded gaze cue to form new word-referent associations and also showed learning in the salience condition. However, their behavior in the salience condition and in the experiments that followed suggests that, rather than basing their judgments of the words' reference on the mere presence or absence of the referent's motion, infants were strongly biased to attend to the consistency with which potential referents moved when a word was heard. (c) 2006 Elsevier Inc. All rights reserved.

Learning associations between places and visual cues without learning to navigate: neither fornix nor entorhinal cortex is required

Rats with fornix transection, or with cytotoxic retrohippocampal lesions that removed entorhinal cortex plus ventral subiculum, performed a task that permits incidental learning about either allocentric (Allo) or egocentric (Ego) spatial cues without the need to navigate by them. Rats learned eight visual discriminations among computer-displayed scenes in a Y-maze, using the constant-negative paradigm. Every discrimination problem included two familiar scenes (constants) and many less familiar scenes (variables). On each trial, the rats chose between a constant and a variable scene, with the choice of the variable rewarded. In six problems, the two constant scenes had correlated spatial properties, either Alto (each constant appeared always in the same maze arm) or Ego (each constant always appeared in a fixed direction from the start arm) or both (Allo + Ego). In two No-Cue (NC) problems, the two constants appeared in randomly determined arms and directions. Intact rats learn problems with an added Allo or Ego cue faster than NC problems; this facilitation provides indirect evidence that they learn the associations between scenes and spatial cues, even though that is not required for problem solution. Fornix and retrohippocampal-lesioned groups learned NC problems at a similar rate to sham-operated controls and showed as much facilitation of learning by added spatial cues as did the controls; therefore, both lesion groups must have encoded the spatial cues and have incidentally learned their associations with particular constant scenes. Similar facilitation was seen in subgroups that had short or long prior experience with the apparatus and task. Therefore, neither major hippocampal input-output system is crucial for learning about allocentric or egocentric cues in this paradigm, which does not require rats to control their choices or navigation directly by spatial cues.

Latent learning in end stage renal disease (ESRD)

Cognitive functions such as attention and memory are known to be impaired in End Stage Renal Disease (ESRD), but the sites of the neural changes underlying these impairments are uncertain. Patients and controls took part in a latent learning task, which had previously shown a dissociation between patients with Parkinson’s disease and those with medial temporal damage. ESRD patients (n=24) and age and education-matched controls (n=24) were randomly assigned to either an exposed or unexposed condition. In Phase 1 of the task, participants learned that a cue (word) on the back of a schematic head predicted that the subsequently seen face would be smiling. For the exposed (but not unexposed) condition, an additional (irrelevant) colour cue was shown during presentation. In Phase 2, a different association, between colour and facial expression, was learned. Instructions were the same for each phase: participants had to predict whether the subsequently viewed face was going to be happy or sad. No difference in error rate between the groups was found in Phase 1, suggesting that patients and controls learned at a similar rate. However, in Phase 2, a significant interaction was found between group and condition, with exposed controls performing significantly worse than unexposed (therefore demonstrating learned irrelevance). In contrast, exposed patients made a similar number of errors to unexposed in Phase 2. The pattern of results in ESRD was different from that previously found in Parkinson’s disease, suggesting a different neural origin.

Cognitive Therapy Abilities in People with Learning Disabilities

Background  There is a need to develop and adapt therapies for use with people with learning disabilities who have mental health problems. Aims  To examine the performance of people with learning disabilities on two cognitive therapy tasks (emotion recognition and discrimination among thoughts, feelings and behaviours). We hypothesized that cognitive therapy task performance would be significantly correlated with IQ and receptive vocabulary, and that providing a visual cue would improve performance. Method  Fifty-nine people with learning disabilities were assessed on the Wechsler Abbreviated Scale of Intelligence (WASI), the British Picture Vocabulary Scale-II (BPVS-II), a test of emotion recognition and a task requiring participants to discriminate among thoughts, feelings and behaviours. In the discrimination task, participants were randomly assigned to a visual cue condition or a no-cue condition. Results  There was considerable variability in performance. Emotion recognition was significantly associated with receptive vocabulary, and discriminating among thoughts, feelings and behaviours was significantly associated with vocabulary and IQ. There was no effect of the cue on the discrimination task. Conclusion  People with learning disabilities with higher IQs and good receptive vocabulary were more likely to be able to identify different emotions and to discriminate among thoughts, feelings and behaviours. This implies that they may more easily understand the cognitive model. Structured ways of simplifying the concepts used in cognitive therapy and methods of socialization and education in the cognitive model are required to aid participation of people with learning disabilities.

Nothing is safe: intolerance of uncertainty is associated with compromised fear extinction learning

Extinction-resistant fear is considered to be a central feature of pathological anxiety. Here we sought to determine if individual differences in Intolerance of Uncertainty (IU), a potential risk factor for anxiety disorders, underlies compromised fear extinction. We tested this hypothesis by recording electrodermal activity in 38 healthy participants during fear acquisition and extinction. We assessed the temporality of fear extinction, by examining early and late extinction learning. During early extinction, low IU was associated with larger skin conductance responses to learned threat vs. safety cues, whereas high IU was associated with skin conductance responding to both threat and safety cues, but no cue discrimination. During late extinction, low IU showed no difference in skin conductance between learned threat and safety cues, whilst high IU predicted continued fear expression to learned threat, indexed by larger skin conductance to threat vs. safety cues. These findings suggest a critical role of uncertainty-based mechanisms in the maintenance of learned fear.

Failure cue priming and impaired cognitive performance—analyses of avoidance motivation as a mediator and fear of failure as a moderator.

The present research investigates whether and how learned symbols for failure reduce task performance. We tested the effect of number priming in two countries with different learning histories for numbers. Priming numbers associated with failure (6 in Germany and 1 in Switzerland) were hypothesized to reduce performance. As expected, in Switzerland, priming with the failure number 1 reduced performance (Study 1), whereas in Germany, priming with the failure number 6 impaired performance in analogy tasks (Study 2). Study 2 additionally analyzed the mechanism and showed that the relationship between failure number priming and performance was mediated by evoked avoidance motivation and that dispositional fear of failure moderated this mediation.

Effects of ocean acidification on learning in coral reef fishes

Ocean acidification has the potential to cause dramatic changes in marine ecosystems. Larval damselfish exposed to concentrations of CO2 predicted to occur in the mid- to late-century show maladaptive responses to predator cues. However, there is considerable variation both within and between species in CO2 effects, whereby some individuals are unaffected at particular CO2 concentrations while others show maladaptive responses to predator odour. Our goal was to test whether learning via chemical or visual information would be impaired by ocean acidification and ultimately, whether learning can mitigate the effects of ocean acidification by restoring the appropriate responses of prey to predators. Using two highly efficient and widespread mechanisms for predator learning, we compared the behaviour of pre-settlement damselfish Pomacentrus amboinensis that were exposed to 440 µatm CO2 (current day levels) or 850 µatm CO2, a concentration predicted to occur in the ocean before the end of this century. We found that, regardless of the method of learning, damselfish exposed to elevated CO2 failed to learn to respond appropriately to a common predator, the dottyback, Pseudochromis fuscus. To determine whether the lack of response was due to a failure in learning or rather a short-term shift in trade-offs preventing the fish from displaying overt antipredator responses, we conditioned 440 or 700 µatm-CO2 fish to learn to recognize a dottyback as a predator using injured conspecific cues, as in Experiment 1. When tested one day post-conditioning, CO2 exposed fish failed to respond to predator odour. When tested 5 days post-conditioning, CO2 exposed fish still failed to show an antipredator response to the dottyback odour, despite the fact that both control and CO2-treated fish responded to a general risk cue (injured conspecific cues). These results indicate that exposure to CO2 may alter the cognitive ability of juvenile fish and render learning ineffective.

Role of Dopamine D2 Receptors in the Consolidation of Amphetamine-Cue Memory in Conditioned Activity in Rats

The neurotransmitter dopamine (DA) plays an essential role in reward-related incentive learning, whereby neutral stimuli gain the ability to elicit approach and other responses. In an incentive learning paradigm called conditioned activity, animals receive a stimulant drug in a specific environment over the course of several days. When then placed in that environment drug-free, they generally display a conditioned hyperactive response. Modulating DA transmission at different time points during the paradigm has been shown to disrupt or enhance conditioning effects. For instance, blocking DA D2 receptors before sessions generally impedes the acquisition of conditioned activity. To date, no studies have examined the role of D2 receptors in the consolidation phase of conditioned activity; this phase occurs immediately after acquisition and involves the stabilization of memories for long-term storage. To investigate this possible role, I trained Wistar rats (N = 108) in the conditioned activity paradigm produced by amphetamine (2.0 mg/kg, intraperitoneally) to examine the effects of the D2 antagonist haloperidol (doses 0.10, 0.25, 0.50, 0.75, 1.0, & 2.0 mg/kg, intraperitoneally) administered 5 min after conditioning sessions. Two positive control groups received haloperidol 1 h before conditioning sessions (doses 1.0 mg/kg and 2.0 mg/kg). The results revealed that post-session haloperidol at all doses tested did not disrupt the consolidation of conditioned activity, while pre-session haloperidol at 2.0 mg/kg prevented acquisition, with the 1.0 mg/kg group trending toward a block. Additionally, post-session haloperidol did not diminish activity during conditioning days, unlike pre-session haloperidol. One possible reason for these findings is that the consolidation phase may have begun earlier than when haloperidol was administered, since the conditioned activity paradigm uses longer learning sessions than those generally used in consolidation studies. Future studies may test if conditioned activity can be achieved with shorter sessions; if so, haloperidol would then be re-tested at an earlier time point. D2 receptor second messenger systems may also be investigated in consolidation. Since drug-related incentive stimuli can evoke cravings in those with drug addiction, a better understanding of the mechanisms of incentive learning may lead to the development of solutions for these individuals.

Implicit learning is not subject to blocking

Explicit (aware) learning has been shown to evidence certain characteristics, such as extinction, blocking, occasion setting, and reliance on context. These characteristics have not been assessed in implicit (unaware) learning. The current study investigated whether implicit learning is subject to blocking. Participants completed a cued reaction time task, where they watched rapid presentations of a random sequence of 8 pairs of shapes, and responded to two target shapes. One target was always preceded by a cue. The experimental group completed a pretraining phase where half the cue, one shape, was followed by the target. Both experimental and control groups completed a training phase where both elements of the cue, two shapes, were followed by the target. Both aware and unaware participants evidenced learning, whereby responding was faster for cued than uncued targets. Aware participants in the experimental group responded faster to targets preceded by the pretrained element than by the other element of the cue. Control and unaware experimental participants were faster to respond to targets preceded by either element of the cue. As blocking was only evident in aware participants, but implicit learning was observed in all participants, it is concluded that implicit learning is not subject to blocking.

Two -way direction of influence in mother-infant interaction as denoted by concurrent reciprocal learning processes

This series of 5 single-subject studies used the operant conditioning paradigm to investigate, within the two-way influence process, how (a) contingent infant attention can reinforce maternal verbal behaviors during a period of mother-infant interaction and under subsequent experimental manipulation. Differential reinforcement was used to determine if it is possible that an infant attending to the mother (denoted by head-turns towards the image of the mother plus eye contact) increases (reinforces) the mother's verbal response (to a cue from the infant) upon which the infant behavior is contingent. There was also (b) an evaluation during the contrived parent-infant interaction for concurrent operant learning of infant vocal behavior via contingent verbal responding (reinforcement) implemented by the mother. Further, it was noted (c) whether or not the mother reported being aware that her responses were influenced by the infant's behavior. Findings showed: the operant conditioning of the maternal verbal behaviors were reinforced by contingent infant attention; and the operant conditioning of infant vocalizations was reinforced by contingent maternal verbal behaviors. No parent reported (1) being aware of the increase in their verbal response reinforced during operant conditioning of parental behavior nor a decrease in those responses during the DRA reversal phase, or (2) noticing a contingency between infant's and mother's response. By binomial 1-tail tests, the verbal-behavior patterns of the 5 mothers were conditioned by infant reinforcement (p < 0.02) and, concurrently, the vocal-response patterns of the 5 infants were conditioned by maternal reinforcement (p < 0.02). A program of systematic empirical research on the determinants of concurrent conditioning within mother-child interaction may provide a way to evaluate the differential effectiveness of interventions aimed at improving parent-child interactions. The work conducted in the present study is one step in this direction. ^

Role of Dopamine D2 Receptors in the Consolidation of Amphetamine-Cue Memory in Conditioned Activity in Rats

The neurotransmitter dopamine (DA) plays an essential role in reward-related incentive learning, whereby neutral stimuli gain the ability to elicit approach and other responses. In an incentive learning paradigm called conditioned activity, animals receive a stimulant drug in a specific environment over the course of several days. When then placed in that environment drug-free, they generally display a conditioned hyperactive response. Modulating DA transmission at different time points during the paradigm has been shown to disrupt or enhance conditioning effects. For instance, blocking DA D2 receptors before sessions generally impedes the acquisition of conditioned activity. To date, no studies have examined the role of D2 receptors in the consolidation phase of conditioned activity; this phase occurs immediately after acquisition and involves the stabilization of memories for long-term storage. To investigate this possible role, I trained Wistar rats (N = 108) in the conditioned activity paradigm produced by amphetamine (2.0 mg/kg, intraperitoneally) to examine the effects of the D2 antagonist haloperidol (doses 0.10, 0.25, 0.50, 0.75, 1.0, & 2.0 mg/kg, intraperitoneally) administered 5 min after conditioning sessions. Two positive control groups received haloperidol 1 h before conditioning sessions (doses 1.0 mg/kg and 2.0 mg/kg). The results revealed that post-session haloperidol at all doses tested did not disrupt the consolidation of conditioned activity, while pre-session haloperidol at 2.0 mg/kg prevented acquisition, with the 1.0 mg/kg group trending toward a block. Additionally, post-session haloperidol did not diminish activity during conditioning days, unlike pre-session haloperidol. One possible reason for these findings is that the consolidation phase may have begun earlier than when haloperidol was administered, since the conditioned activity paradigm uses longer learning sessions than those generally used in consolidation studies. Future studies may test if conditioned activity can be achieved with shorter sessions; if so, haloperidol would then be re-tested at an earlier time point. D2 receptor second messenger systems may also be investigated in consolidation. Since drug-related incentive stimuli can evoke cravings in those with drug addiction, a better understanding of the mechanisms of incentive learning may lead to the development of solutions for these individuals.

Child Safety Signal Learning in the Context of Parental Worry, Overprotection, PTSD, and Depression

Thesis (Ph.D.)--University of Washington, 2016-08

A study of the role of a technology-enhanced learning implementation group in mediating an institutional VLE minimum standards policy

Recent years have seen a focus on responding to student expectations in higher education. As a result, a number of technology-enhanced learning (TEL) policies have stipulated a requirement for a minimum virtual learning environment (VLE) standard to provide a consistent student experience. This paper offers insight into an under-researched area of such a VLE standard policy development using a case study of one university. With reference to the implementation staircase model, this study takes cue from the view that an institutional VLE template can affect lower levels directly, sidestepping the chain in the implementation staircase. The Group's activity whose remit is to design and develop a VLE template, therefore, becomes significant. The study, drawing on activity theory, explores the mediating role of such a Group. Factors of success and sources of tension are analysed to understand the interaction between the individuals and the collective agency of Group members. The paper identifies implications to practice for similar TEL development projects. Success factors identified demonstrated the importance of good project management principles, establishing clear rules and division of labour for TEL development groups. One key finding is that Group members are needed to draw on both different and shared mediating artefacts, supporting the conclusion that the nature of the group's composition and the situated expertise of its members are crucial for project success. The paper's theoretical contribution is an enhanced representation of a TEL policy implementation staircase.

Ecological Literacy And Beyond: Problem-based Learning For Future Professionals.

Ecological science contributes to solving a broad range of environmental problems. However, lack of ecological literacy in practice often limits application of this knowledge. In this paper, we highlight a critical but often overlooked demand on ecological literacy: to enable professionals of various careers to apply scientific knowledge when faced with environmental problems. Current university courses on ecology often fail to persuade students that ecological science provides important tools for environmental problem solving. We propose problem-based learning to improve the understanding of ecological science and its usefulness for real-world environmental issues that professionals in careers as diverse as engineering, public health, architecture, social sciences, or management will address. Courses should set clear learning objectives for cognitive skills they expect students to acquire. Thus, professionals in different fields will be enabled to improve environmental decision-making processes and to participate effectively in multidisciplinary work groups charged with tackling environmental issues.