The past twenty years have seen a successful formalization of the idea that perception is a form of probabilistic inference. Bayesian Decision Theory (BDT) provides a neat mathematical framework for describing how an ideal observer and actor should interpret incoming sensory stimuli and act in the face of uncertainty. The predictions of BDT, however, crucially depend on the observer’s internal models, represented in the Bayesian framework by priors, likelihoods, and the loss function. Arguably, only in the simplest scenarios (e.g., with a few Gaussian variables) we can expect a real observer’s internal representations to perfectly match the true statistics of the task at hand, and to conform to exact Bayesian computations, but how humans systematically deviate from BDT in more complex cases is yet to be understood. In this thesis we theoretically and experimentally investigate how people represent and perform probabilistic inference with complex (beyond Gaussian) one-dimensional distributions of stimuli in the context of sensorimotor decision making. The goal is to reconstruct the observers’ internal representations and details of their decision-making process from the behavioural data – by employing Bayesian inference to uncover properties of a system, the ideal observer, that is believed to perform Bayesian inference itself. This “inverse problem” is not unique: in principle, distinct Bayesian observer models can produce very similar behaviours. We circumvented this issue by means of experimental constraints and independent validation of the results. To understand how people represent complex distributions of stimuli in the specific domain of time perception, we conducted a series of psychophysical experiments where participants were asked to reproduce the time interval between a mouse click and a flash, drawn from a session-dependent distribution of intervals. We found that participants could learn smooth approximations of the non-Gaussian experimental distributions, but seemed to have trouble with learning some complex statistical features such as bimodality. To investigate whether this difficulty arose from learning complex distributions or computing with them, we conducted a target estimation experiment in which “priors” where explicitly displayed on screen and therefore did not need to be learnt. Lack of difference in performance between the Gaussian and bimodal conditions in this task suggests that acquiring a bimodal prior, rather than computing with it, is the major difficulty. Model comparison on a large number of Bayesian observer models, representing different assumptions about the noise sources and details of the decision process, revealed a further source of variability in decision making that was modelled as a “stochastic posterior”. Finally, prompted by a secondary finding of the previous experiment, we tested the effect of decision uncertainty on the capacity of the participants to correct for added perturbations in the visual feedback in a centre of mass estimation task. Participants almost completely compensated for the injected error in low uncertainty trials, but only partially so in the high uncertainty ones, even when allowed sufficient time to adjust their response. Surprisingly, though, their overall performance was not significantly affected. This finding is consistent with the behaviour of a Bayesian observer with an additional term in the loss function that represents “effort” – a component of optimal control usually thought to be negligible in sensorimotor estimation tasks. Together, these studies provide new insight into the capacity and limitations people have in learning and performing probabilistic inference with distributions beyond Gaussian. This work also introduces several tools and techniques that can help in the systematic exploration of suboptimal behaviour. Developing a language to describe suboptimality, mismatching representations and approximate inference, as opposed to optimality and exact inference, is a fundamental step to link behavioural studies to actual neural computations.

Existing computational cognitive models of spatial memory often neglect difficulties posed by the real world, such as sensory noise, uncertainty, and high spatial complexity. On the other hand, robotics is unconcerned with understanding biological cognition. This thesis takes an interdisciplinary approach towards developing cognitively plausible spatial memory models able to function in realistic environments, despite sensory noise and spatial complexity. We hypothesized that Bayesian localization and error correction accounts for how brains might maintain accurate location estimates, despite sensory errors. We argued that these mechanisms are psychologically plausible (producing human-like behaviour) as well as neurally plausible (implementable in brains). To support our hypotheses, we reported modelling results of neural recordings from rats (acquired outside this PhD), constituting the first evidence for Bayesian inference in neurons representing spatial location, as well as modelling human behaviour data. In addition to dealing with uncertainty, spatial representations have to be stored and used efficiently in realistic environments, by using structured representations such as hierarchies (which facilitate efficient retrieval and route planning). Evidence suggests that human spatial memories are structured hierarchically, but the process responsible for these structures has not been known. We investigated features influencing them using data from experiments in real-world and virtual reality environments, and proposed a computational model able to predict them in advance (based on clustering in psychological space). We have extended a general cognitive architecture, LIDA (Learning Intelligent Distribution Agent), by these probabilistic models of how brains might estimate, correct, and structure representations of spatial locations. We demonstrated the ability of the resulting model to deal with the challenges of realistic environments by running it in high-fidelity robotic simulations, modelled after participants' actual cities. Our results show that the model can deal with noise, uncertainty and complexity, and that it can reproduce the spatial accuracies of human participants.

Cotutelle entre l’université de Montréal et d’Aix-Marseille / Cette thèse vise à comprendre les fondements et les fonctions des calculs probabilistes impliqués dans les processus visuels. Nous nous appuyons sur une double stratégie, qui implique le développement de modèles dans le cadre du codage prédictif selon le principe de l'énergie libre. Ces modèles servent à définir des hypothèses claires sur la fonction neuronale, qui sont testées à l'aide d'enregistrements extracellulaires du cortex visuel primaire. Cette région du cerveau est principalement impliquée dans les calculs sur les unités élémentaires des entrées visuelles naturelles, sous la forme de distributions d'orientations. Ces distributions probabilistes, par nature, reposent sur le traitement de la moyenne et de la variance d'une entrée visuelle. Alors que les premières ont fait l'objet d'un examen neurobiologique approfondi, les secondes ont été largement négligées. Cette thèse vise à combler cette lacune. Nous avançons l'idée que la connectivité récurrente intracorticale est parfaitement adaptée au traitement d'une telle variance d'entrées, et nos contributions à cette idée sont multiples. (1) Nous fournissons tout d'abord un examen informatique de la structure d'orientation des images naturelles et des stratégies d'encodage neuronal associées. Un modèle empirique clairsemé montre que le code neuronal optimal pour représenter les images naturelles s'appuie sur la variance de l'orientation pour améliorer l'efficacité, la performance et la résilience. (2) Cela ouvre la voie à une étude expérimentale des réponses neurales dans le cortex visuel primaire du chat à des stimuli multivariés. Nous découvrons de nouveaux types de neurones fonctionnels, dépendants de la couche corticale, qui peuvent être liés à la connectivité récurrente. (3) Nous démontrons que ce traitement de la variance peut être compris comme un graphe dynamique pondéré conditionné par la variance sensorielle, en utilisant des enregistrements du cortex visuel primaire du macaque. (4) Enfin, nous soutenons l'existence de calculs de variance (prédictifs) en dehors du cortex visuel primaire, par l'intermédiaire du noyau pulvinar du thalamus. Cela ouvre la voie à des études sur les calculs de variance en tant que calculs neuronaux génériques soutenus par la récurrence dans l'ensemble du cortex. / This thesis aims to understand the foundations and functions of the probabilistic computations involved in visual processes. We leverage a two-fold strategy, which involves the development of models within the framework of predictive coding under the free energy principle. These models serve to define clear hypotheses of neuronal function, which are tested using extracellular recordings of the primary visual cortex. This brain region is predominantly involved in computations on the elementary units of natural visual inputs, in the form of distributions of oriented edges. These probabilistic distributions, by nature, rely on processing both the mean and variance of a visual input. While the former have undergone extensive neurobiological scrutiny, the latter have been largely overlooked. This thesis aims to bridge this knowledge gap. We put forward the notion that intracortical recurrent connectivity is optimally suited for processing such variance of inputs, and our contributions to this idea are multi-faceted. (1) We first provide a computational examination of the orientation structure of natural images and associated neural encoding strategies. An empirical sparse model shows that the optimal neural code for representing natural images relies on orientation variance for increased efficiency, performance, and resilience. (2) This paves the way for an experimental investigation of neural responses in the cat's primary visual cortex to multivariate stimuli. We uncover novel, cortical-layer-dependent, functional neuronal types that can be linked to recurrent connectivity. (3) We demonstrate that this variance processing can be understood as a dynamical weighted graph conditioned on sensory variance, using macaque primary visual cortex recordings. (4) Finally, we argue for the existence of (predictive) variance computations outside the primary visual cortex, through the Pulvinar nucleus of the thalamus. This paves the way for studies on variance computations as generic weighting of neural computations, supported by recurrence throughout the entire cortex.

La perception de notre environnement repose sur les informations sensorielles reçues, mais aussi sur nos a priori. Dans le cadre Bayésien, ces a priori capturent les régularités de notre environnement et sont essentiels pour inférer les causes de nos sensations. Récemment, les théories du cerveau Bayésien ont été appliquées à l'autisme pour tenter d'en expliquer les symptômes. Les troubles du spectre de l'autisme (TSA) sont caractérisés par des difficultés de compréhension des interactions sociales, par des comportements restreints et répétitifs, et par une perception sensorielle atypique.Cette thèse vise à caractériser l'inférence et l'apprentissage perceptifs dans les TSA, en étudiant la sensorialité et la construction d'a priori. Nous avons utilisé des tests comportementaux, des modèles computationnels, des questionnaires, de l'imagerie fonctionnelle et de la spectroscopie par résonnance magnétique chez des adultes avec ou sans TSA. La définition des profils sensoriels de personnes avec des hauts quotients autistiques a été affinée grâce à un questionnaire dont nous avons validé la traduction française. En explorant les stratégies d'apprentissage perceptif, nous avons ensuite montré que les personnes avec TSA étaient moins enclines à spontanément utiliser une mode d'apprentissage permettant de généraliser. L'étude de la construction implicite des a priori a montré que les personnes avec TSA étaient capables d'apprendre un a priori, mais l'ajustaient difficilement suite à un changement de contexte. Enfin, l'étude des corrélats neurophysiologiques de l'inférence perceptive a révélé un réseau cérébral et une neuromodulation différents dans les TSA.L'ensemble de ces résultats met en lumière une perception atypique dans les TSA, marquée par un apprentissage et une pondération anormale des a priori. Une approche Bayésienne des TSA pourrait améliorer leur caractérisation, diagnostics et prises en charge / How we perceive our environment relies both on sensory information and on our priors or expectations. Within the Baysian framework, these priors capture the underlying statistical regularities of our environment and allow inferring sensation causes. Recently, Bayesian brain theories suggested that autistic symptoms could arise from an atypical weighting of sensory information and priors. Autism spectrum disorders (ASD) is characterized defined by difficulties in social interactions, by restricted and repetitive patterns of behaviors, and by an atypical sensory perception.This thesis aims at characterizing perceptual inference and learning in ASD, and studies sensory sensitivity and prior learning. This was investigated using behavioral tasks, computational models, questionnaires, functional magnetic resonance imaging and magnetic resonance spectroscopy in adults with or without ASD. Sensory profiles in people with high autism spectrum quotients were first refined, using a questionnaire that we validated in French. The study of perceptual learning strategies then revealed that subjects with ASD were less inclined to spontaneously use a learning style enabling generalization. The implicit learning of priors was explored and showed that subjects with ASD were able to build up a prior but had difficulties adjusting it in changing contexts. Finally, the investigation of the neurophysiological correlates and molecular underpinnings of a similar task showed that perceptual decisions biased by priors relied on a distinct neural network in ASD, and was not related to the same modulation by the glutamate/GABA ratio.The overall results shed light on an atypical learning and weighting of priors in ASD, resulting in an abnormal perceptual inference. A Bayesian approach could help characterizing ASD and could contribute to ASD diagnosis and care

Cette thèse porte sur le codage prédictif comme principe général pour la perception et vise à en étayer les mécanismes computationnels et neurophysiologiques dans la modalité auditive. Ce codage repose sur des erreurs de prédictions se propageant dans une hiérarchie, et qui pourraient se refléter dans des réponses cérébrales au changement (ou déviance) telles que la Négativité de discordance (mismatch negativity, MMN). Nous avons manipulé la prédictibilité de sons déviants et utilisé des approches de modélisation computationnelle et dynamique causale (DCM) appliquées à des enregistrements électrophysiologiques (EEG, MEG) simultanés.Une modulation des réponses à la déviance par la prédictibilité a été observée, permettant d'établir un lien avec les erreurs de prédictions. Cet effet implique un apprentissage implicite des régularités acoustiques, dont l'influence sur le traitement auditif a pu être caractérisée par notre approche de modélisation. Sur le plan computationnel, un apprentissage a été mis en évidence au cours de ce traitement auditif, reposant sur une fenêtre d'intégration temporelle dont la taille varie avec la prédictibilité des déviants. Cet effet pourrait également moduler la connectivité synaptique sous-tendant le traitement auditif, comme le suggère l'analyse DCM.Nos résultats mettent en évidence la mise en œuvre d'un apprentissage perceptif au sein d'une hiérarchie auditive soumis à une modulation par la prédictibilité du contexte acoustique, conformément aux prédictions du codage prédictif. Ils soulignent également l'intérêt de ce cadre théorique pour émettre et tester expérimentalement des hypothèses mécanistiques précises / This thesis aims at testing the predictive coding account of auditory perception. This framework rests on precision-weighted prediction errors elicited by unexpected sounds that propagate along a hierarchical organization in order to maintain the brain adapted to a varying acoustic environment. Using the mismatch negativity (MMN), a brain response to unexpected stimuli (deviants) that could reflect such errors, we could address the computational and neurophysiological underpinnings of predictive coding. Precisely, we manipulated the predictability of deviants and applied computational learning models and dynamic causal models (DCM) to electrophysiological responses (EEG, MEG) measured simultaneously. Deviant predictability was found to modulate deviance responses, a result supporting their interpretation as prediction errors. Such effect might involve the (high-level) implicit learning of sound sequence regularities that would in turn influence auditory processing in lower hierarchical levels. Computational modeling revealed the perceptual learning of sounds, resting on temporal integration exhibiting differences induced by our predictability manipulation. In addition, DCM analysis indicated predictability changes in the synaptic connectivity established by deviance processing. These results conform predictive coding predictions regarding both deviance processing and its modulation by deviant predictability and strongly support perceptual learning of auditory regularities achieved within an auditory hierarchy. Our findings also highlight the power of this mechanistic framework to elaborate and test new hypothesis enabling to improve our understanding of auditory processing