Yes / To enable effective interaction with the environment, the brain combines noisy sensory information with expectations based on prior experience. There is ample evidence showing that humans can learn statistical regularities in sensory input and exploit this knowledge to improve perceptual decisions and actions. However, fundamental questions remain regarding how priors are learned and how they generalize to different sensory and behavioral contexts. In principle, maintaining a large set of highly specific priors may be inefficient and restrict the speed at which expectations can be formed and updated in response to changes in the environment. However, priors formed by generalizing across varying contexts may not be accurate. Here, we exploit rapidly induced contextual biases in duration reproduction to reveal how these competing demands are resolved during the early stages of prior acquisition. We show that observers initially form a single prior by generalizing across duration distributions coupled with distinct sensory signals. In contrast, they form multiple priors if distributions are coupled with distinct motor outputs. Together, our findings suggest that rapid prior acquisition is facilitated by generalization across experiences of different sensory inputs but organized according to how that sensory information is acted on.

Despite a rise in the use of “learning by doing” pedagogical methods in praxis, little is known as to how the brain benefits from these methods. Learning by doing strategies that utilize complementary information (“enrichment”) such as gestures have been shown to optimize learning outcomes in several domains including foreign language (L2) training. Here we tested the hypothesis that behavioral benefits of gesture-based enrichment are critically supported by integrity of the biological motion visual cortices (bmSTS). Prior functional neuroimaging work has implicated the visual motion cortices in L2 translation following sensorimotor-enriched training; the current study is the first to investigate the causal relevance of these structures in learning by doing contexts. Using neuronavigated transcranial magnetic stimulation and a gesture-enriched L2 vocabulary learning paradigm, we found that the bmSTS causally contributed to behavioral benefits of gesture-enriched learning. Visual motion cortex integrity benefitted both short- and long-term learning outcomes, as well as the learning of concrete and abstract words. These results adjudicate between opposing predictions of two neuroscientific learning theories: While reactivation-based theories predict no functional role of specialized sensory cortices in vocabulary learning outcomes, the current study supports the predictive coding theory view that these cortices precipitate sensorimotor-based learning benefits.:I. Abkürzungsverzeichnis II. Abbildungsverzeichnis III. Einleitung 1. Fremdsprachenlernen 1.1 Sensorische Modalitätsvergleiche 1.2 Sensomotorisches Lernen 2. Lerntheorien 2.1 Theorie des prädiktiven Kodierens 2.2 Theorie des prädiktiven Kodierens für multisensorisches Lernen 3. Sulcus temporalis superior für biologische Bewegung 4. Transkranielle Magnetstimulation 4.1 Passagere Funktionsinhibition mittels transkranieller Magnetstimulation IV. Ableitung der Rationale V. Publikationsmanuskript VI. Zusammenfassung VII. Literaturverzeichnis VIII. Appendix A. Abbildungen B. Ergänzendes Material der Publikation C. Darstellung des eigenen Beitrags D. Erklärung über die eigenständige Abfassung der Arbeit E. Lebenslauf F. Publikationen G. Danksagung

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.

The fields of biologically inspired artificial intelligence, neuroscience, and psychology have had exciting influences on each other over the past decades. Especially recently, with the increased popularity and success of artificial neural networks (ANNs), ANNs have enjoyed frequent use as models for brain function. However, there are still many disparities between the implementation, algorithms, and learning environment used for deep learning and those employed by the brain, which is reflected in their differing abilities. I first briefly introduce ANNs and survey the differences and similarities between them and the brain. I then make a case for designing the learning environment of ANNs to be more similar to that in which brains learn, namely by allowing them to actively interact with the world and decreasing the amount of external supervision. To implement this sensorimotor learning in an artificial agent, I use deep reinforcement learning, which I will also briefly introduce and compare to learning in the brain. In the research presented in this dissertation, I focus on testing the hypothesis that the learning environment matters and that learning in an embodied way leads to acquiring different representations of the world. We first tested this on human subjects, comparing spatial knowledge acquisition in virtual reality to learning from an interactive map. The corresponding two publications are complemented by a methods paper describing eye tracking in virtual reality as a helpful tool in this type of research. After demonstrating that subjects do indeed learn different spatial knowledge in the two conditions, we test whether this transfers to artificial agents. Two further publications show that an ANN learning through interaction learns significantly different representations of the sensory input than ANNs that learn without interaction. We also demonstrate that through end-to-end sensorimotor learning, an ANN can learn visually-guided motor control and navigation behavior in a complex 3D maze environment without any external supervision using curiosity as an intrinsic reward signal. The learned representations are sparse, encode meaningful, action-oriented information about the environment, and can perform few-shot object recognition despite not knowing any labeled data beforehand. Overall, I make a case for increasing the realism of the computational tasks ANNs need to solve (largely self-supervised, sensorimotor learning) to improve some of their shortcomings and make them better models of the brain.

Dans cette thèse, nous nous intéressons à l'étude d'un modèle dédié à l'exploration et à la structuration d'espaces d'apprentissage sensorimoteur pour des systèmes artificiels. Nous appuyons notre démarche sur les notions de corps et de développement propre, auxquelles se greffe un troisième processus dit motivationnel. Cette forme de curiosité artificielle se base sur le progrès en compétence et repose ainsi sur les contraintes physiques naturelles directement issues de l'encorporation de l'agent. L'objectif de la motivation est de réguler un développement à long terme, dédié à l'apprentissage de nouvelles compétences non prévues par le concepteur. Nous inscrivons nos travaux dans la continuité de l'approche du babillage sensorimoteur dans l'espace des buts, qui consiste à déterminer un ensemble de techniques permettant à un agent de générer, selon une métrique d'intérêt, une configuration sensorielle qu'il va essayer d'atteindre par des actions motrices. Nos contributions viennent améliorer et complexifier un modèle motivationnel existant, en proposant des alternatives au processus de structuration de l'espace d'exploration. Certaines de ces propositions théoriques ont été validées expérimentalement grâce à la plateforme FIMO, que nous avons développée dans cette optique, et qui est disponible en ligne. / In this thesis, we study a motivational model for artificial systems, which aims at exploring and structuring sensorimotor learning spaces. Our approach relies on some essential notions, including the body, the development, and the motivation. This particular kind of artificial curiosity is based on the competence or learning progress, and thus depends on the physical natural constraints originating from the agent's embodiment. We follow the Goal-Babbling Exploration approach which consists in determining a set of techniques allowing an agent to self-generate goals, i.e. sensory configurations, and try to reach them using motor actions. Our contributions improve the SAGG-RIAC motivational model, by proposing alternative ways of structuring the exploration of the goal space. Some of our contributions have been experimentally validated within the FIMO framework we developed to this purpose.

Si la perception de notre environnement paraît simple et naturelle, les attributs visuels composant les différents objets sont traités dans des espaces temps qui leur sont propres en engageant des voies neuronales distinctes. La question du liage des informations sensorielles en un percept unique est alors posée. Dans le cadre de cette thèse, la question du liage sensoriel a été étudiée au regard des actes moteurs volontaires. En situation d'observation passive, le point de simultanéité subjective (PSS) montre qu'un changement de couleur doit se produire 40 ms avant un changement de position pour donner lieu à une perception synchrone des deux événements. L'exécution d'un mouvement d'atteinte manuelle réduit significativement ce PSS (-3.3 ms) uniquement lorsque le délai et l'écart spatial entre la fin de l'action motrice et les changements environnementaux n'excède pas 250 ms (-0,6 ms) et 2 cm ( 3,8 ms). Cependant, dans le cas où une situation d'apprentissage est induite par la présence de nouvelles contingences sensori-motrices, le liage sensoriel par l'action peut être observé pour des intervalles de temps plus grands (jusqu'à 1000 ms). Ce liage sensoriel par l'action, mis en évidence pour la première fois, serait sous-tendu par les mécanismes prédictifs associés aux modèles internes. Une dernière étude a révélé l'intérêt diagnostique de l'étude des relations perception/action au cours du vieillissement. Des profils de réponse spécifique ont été observés chez des patients de type Alzheimer lors de tâches testant les relations perception/action. Ces études apportent des arguments en faveur d'un rôle important de l'action dans la perception et la cognition. / If the perception of our environment seems easy and natural, the visual attributes of various objects are processed in space-time which are their own by engaging distinct neural pathways. The issue of binding of sensory information into a single percept is then arise. In this thesis, the issus of sensory bending was studied in relation to voluntary motor action. In position of passive observation, the point of subjective simultaneity (PSS) shows that a color change had to occur 40 ms before a change of position to give rise to a perception of two synchronous events. Performing a manual reaching movement reduced significantly the pSS (-3.3 ms) only when the time and space gap between the end of the motor action and environmental change does not exceed 250 ms (-0.6 ms) and 2 cm (3.8 ms). However, if a learning situation is induced by the presence of new sensorimotor contingencies, sensory binding by action can be observed for larger time intervals (up to 1000ms). This sensory binding by action, demonstrated for the first time, would be underpinned by predictive mechanisms associated with internal models. A last study showed the diagnostic interest of the study of relation perception/action during aging. Specific profiles of responses were observed in patients with Alzheimer on task testing the relations perception/action. These studies provide arguments for an important role of action in perception and cognition.

Observative prospektive Kohortenstudie mit 22 Patienten über 3 Jahre. Wir untersuchten, ob die intensive Schulung der Diskrimination des Hüftgelenks von Beckenbewegungen Auswirkungen auf die Harmonisierung des Gangbilds in der postoperativen Rehabilitation habe. Es wurde anhand der videogestützten Ganganalyse auf dem Laufband ein Score der Gangharmonie gebildet, der neben dem Gang als Ganzkörperbewegung auch koordinative Faktoren wie den Bewegungsfluß berücksichtigt. Faktoren wie Schmerz und Erfahrung in übenden Verfahren wurden in einer Regressionsanalyse auf ihre Assoziation mit der Gangharmonie überprüft. Modelle des Bewegungslernens wurden auf langfristige Strategien der Rehabilitation angewendet. / Prospective observative cohort study over 3 years. A cohort of people with OA of the hip due to CDH was followed up over 3 years. The discrimination of hip movements from pelvis movements was measured and compared with the gait harmony. A Gait Harmony Score reflecting total body movement and coordinative factors has been developed and validated. In a regression analysis we measured the association between confounding variables like preoperative pain and experience in exercise with gait harmony and hip proprioception. Models of motor learning in rehabilitation have been applied to the postoperative development of the gait in order to establish models for longterm rehabilitation strategies.

Human sensorimotor control is remarkably adept at utilising contextual information to learn and recall systematic sensorimotor transformations. Here, we investigate the motor representations that underlie such learning, and examine how motor memories acquired based on different contextual information interact. Using a novel three-dimensional robotic manipulandum, the 3BOT, we examined the spatial transfer of learning across various movement directions in a 3D environment, while human subjects performed reaching movements under velocity-dependent force field. The obtained pattern of generalisation suggested that the representation of dynamic learning was most likely defined in a target-based, rather than an extrinsic, coordinate system. We further examined how motor memories interact when subjects adapt to force fields applied in orthogonal dimensions. We found that, unlike opposing fields, learning two spatially orthogonal force fields led to the formation of separate motor memories, which neither interfered with nor facilitated each other. Moreover, we demonstrated a novel, more general aspect of the spontaneous recovery phenomenon using a two-dimensional force field task: when subjects learned two orthogonal force fields consecutively, in the following phase of clamped error feedback, the expression of adaptation spontaneously rotated from the direction of the second force field, towards the direction of the first force field. Finally, we examined the interaction of sensorimotor memories formed based on separate contextual information. Subjects performed reciprocating reaching and object manipulation tasks under two alternating contexts (movement directions), while we manipulated the dynamics of the task in each context separately. The results suggested that separate motor memories were formed for the dynamics of the task in different contexts, and that these motor memories interacted by sharing error signals to enhance learning. Importantly, the extent of interaction was not fixed between the context-dependent motor memories, but adaptively changed according to the task dynamics to potentially improve overall performance. Together, our experimental and theoretical results add to the understanding of mechanisms that underlie sensorimotor learning, and the way these mechanisms interact under various tasks and different dynamics.

Le contrôle moteur a traditionnellement été étudié séparément des autres processus cognitifs qui sous-tendent la parole, dans la lignée de théories de la cognition présentant le cerveau comme un ensemble de modules relativement indépendants (Fodor & Pylyshyn, 2007). Cependant les recherches autour de la cognition incarnée (Varela, Thompson, & Rosch, 1991) et située (Barsalou, 2008), ainsi que des systèmes dynamiques (Smith & Thelen, 2003), menées ces trois dernières décennies soulignent que la cognition ne peut pas être considérée séparément d’un corps et de son environnement. Ce cadre constitue une source d’inspiration pour cette thèse et une motivation pour étudier les processus sensorimoteurs de la parole en lien avec les autres processus cognitifs.La parole peut-être décomposée en séquence d’unités linguistiques structurées sous forme d’une hiérarchie. Nous soutenons que ces unités sont ancrées dans des représentations sensorimotrices, associant une structure linguistique avec des informations perceptives et motrices. Ces unités correspondent-elles à des mots ? Des syllabes ? Des phonèmes ? Pour sonder les représentations assurant l’articulation de la parole, nous proposons d’utiliser un paradigme d’apprentissage auditorimoteur basé sur la perturbation du retour auditif (Caudrelier & Rochet-Capellan, accepté). Ce paradigme permet de modifier chez un locuteur des représentations sensorimotrices spécifiques, les représentations qui sous-tendent la production d’un item d’entrainement, par exemple un mot. Nous faisons ainsi l’hypothèse que si cette modification affecte la prononciation d’un autre mot, cela veut dire que la production de ce mot s’appuie sur une partie de ces représentations. Ainsi, l’observation du transfert d’apprentissage permet de révéler la structure de représentations qui assurent la production de parole.Une première étude chez l’adulte montre que le transfert d’apprentissage auditorimoteur a lieu à la fois aux niveaux du phonème, de la syllabe et du mot (Caudrelier, Schwartz, Perrier, Gerber, & Rochet-Capellan, 2018). Ces observations suggèrent que ces unités co-contribuent à l’articulation de la parole chez l’adulte. Les résultats sont mis en perspective par rapport aux théories et modèles de production de parole. Une 2ème expérience suggère que la modalité de présentation du stimulus (un mot à lire ou une image à dénommer) peut influencer le transfert d’apprentissage auditorimoteur (Caudrelier, Perrier, Schwartz, & Rochet-Capellan, 2018). Une 3ème étude chez des enfants de 4-5 ans et de 7-8 ans montre que les représentations du phonème émergent avant l’acquisition de la lecture (Caudrelier et al., en révision). De plus, un lien entre adaptation à la perturbation auditive et conscience phonologique est mis en évidence dans les deux groupes d’âge. Le potentiel caractère prédictif ou causal de ce lien est discuté.En conclusion, cette thèse exploite un outil original et productif pour explorer les représentations de la parole et étudier leur développement. Ce travail pourrait avoir des implications cliniques, pour la rééducation de la parole, et pour la dyslexie développementale. Il met en évidence des liens entre les niveaux sensorimoteurs, linguistiques et contextuels qui questionnent la nature des représentations qui sous-tendent la parole. / Speech motor control has traditionally been studied apart from other cognitive processes underlying speech production, since first cognitive theories presented the brain as a set of relatively independent modules (Fodor & Pylyshyn, 2007), taken apart from the body. However developments in embodied cognition (Varela, Thompson, & Rosch, 1991), grounded cognition (Barsalou, 2008) and dynamic systems (Smith & Thelen, 2003) occurred in the last three decades underline that cognition cannot be considered separately from a body and its environment. These frameworks constitute an inspiration for this thesis and a motivation to study motor control and sensorimotor processes in relation to other cognitive processes. Whether linguistic structures are grounded in sensorimotor processes will be an underlying question.A spoken message can be decomposed into sequences of linguistic units hierarchically structured. We argue that these speech units are grounded in sensorimotor representations, associating linguistic structures with auditory and motor information. Do these units correspond to words? Syllables? Phonemes? To probe the building blocks of speech production, we propose to use a paradigm of auditory-motor learning based on auditory feedback perturbation (Caudrelier & Rochet-Capellan, in press). This paradigm actually enables to change specific internal sensorimotor representations in speakers. Adaptation induces updating sensorimotor representations underlying the production of the training item. We assume that if this change affects the pronunciation of another word, it means that this word uses some of these updated representations. Thus, transfer patterns may reveal the structure of representations at stake.A first study in adults shows that transfer of auditory-motor learning occurs at word, syllable, and phoneme levels in parallel (Caudrelier, Schwartz, Perrier, Gerber, & Rochet-Capellan, 2018). These observations suggest that all these units may co-contribute to the organization of speech articulation in adult speakers. Experimental results are discussed in the light of existing theories and models of speech production. A second experiment suggests that whether a speaker reads a word aloud or names a picture may have an influence on the transfer of auditory-motor learning (Caudrelier, Perrier, Schwartz, & Rochet-Capellan, 2018). A third study in 4- to 5-year-old and 7- to 8 year-old children investigates whether phoneme sensorimotor representations may emerge during reading acquisition, or prior to it (Caudrelier et al., in revision). The observed transfer patterns suggest that phoneme representations emerge before reading acquisition, as a consequence of speech experience. Moreover, we found a relationship between adaptation to auditory perturbation and phonological awareness scores in both age groups. This suggests a link between sensorimotor representations and more explicit phonological representations. The potential causal or predictive nature of this link is discussed.Overall, this work exploits an original and fruitful tool to probe speech representations and study their development. It may have clinical implications with regards to speech rehabilitation, as well as developmental dyslexia. It also highlights connections between speech sensorimotor level and higher linguistic and contextual levels that further question the nature of speech representations.

No description available.