Search

International audience; This paper presents a new approach for learning transition function in state representation learning (SRL) for control. While state-of-the-art methods use different deterministic neural networks to learn forward and inverse state transition functions independently with auto-supervised learning, we introduce a bidirectional stochastic model to learn both transition functions. We aim at using the uncertainty of the model on its predictions as an intrinsic motivation for exploration to enhance the representation learning. More, using the same model to learn both transition functions allows sharing the parameters, which can reduce their number and should increase the embedding quality of the representation. We use a factored restricted Boltzmann machine (fRBM) based model, enhanced with dedicated structure for learning system dynamics and transitions with shared parameters. The presented work focuses on building the structure of the bidirectional transition model for unsupervised learning. Our fRBM structure is directly inspired from physics interactions between inputs and outputs in reinforcement learning framework. We compare different training algorithms for learning the model that must be able to predict observable random variables to be used in SRL framework. Our structure is not restricted to any type of observable, nevertheless in this paper we focus on learning dynamics from the OpenAI Gym environment Swinging Pendulum. We show that the proposed structure is able to learn bidirectional transition function and performs well in prediction task.

We propose a participatory science outreach approach allowing us to co-construct with our audiences resources aimed at understanding and demystifying the most disruptive results obtained regarding human brain by the conjunction of computer science, applied mathematics and neuroscience (computational neuroscience). The context is that of science and technology with a heavy societal impact, for which there is a strong need to allow everyone to build models of representation of these results and to forge an enlightened citizen's vision on these subjects.We rely here on our experience in sharing scientific culture on these subjects and our ability to create large diffusion content and resources, easy to appropriate and to operate.We propose to discover the models of the cerebral functions at the origin of our sensorimotor and vital cognitive behaviors (instinctive and motivated behavior, selection of embodied action, emotional decision-making or not, sites of self-awareness, etc. ) through :- a course of evolving content each time giving minimal key ideas on these subjects, also showing the simple use of mathematical concepts,- a Web-application (3D visualization of the brain in synergy with multi-media content and explanatory texts) with the possibility of interacting with the content. e.g., quizzes.The implementation is a free and open code, easily reusable by anyone with basic computing skills.This is also in itself a tool for learning the code, in addition to the acquisition of skills in integrative neuroscience, and it is a lever for co-creation.; On propose la mise en place d’une démarche de médiation scientifique participative pour permettre de co-construire avec nos publics des ressources visant à comprendre et démystifier les résultats les plus disruptifs concernant le cerveau humain obtenus par la conjonction de l’informatique, mathématiques appliquées et des neurosciences (neurosciences computationnelles).Le contexte est celui de sciences et technologies à lourd impact sociétal avec un besoin fort de permettre à chacune et chacun de se construire des modèles de représentation de ces résultats et de se forger une vision citoyenne éclairée sur ces sujets. On s’appuie ici sur notre expérience en matière de partage de culture scientifique sur ces sujets et notre capacité à créer des contenus et des ressources, à forte diffusion, faciles à s’approprier et à faire fonctionner.On propose de découvrir les modèles des fonctions cérébrales à l’origine de nos comportements sensori-moteurs et cognitifs vitaux (comportement instinctif et motivé, sélection de l’action incarnée, prise de décision émotionnelle ou non, siège de la conscience de soi, …) à travers :- un parcours de contenus évolutifs donnant à chaque fois des idées clés minimales sur ces sujets, en montrant aussi l’utilisation simple de notions mathématiques, - une Web-application (visualisation 3D du cerveau en synergie avec des contenus multi-médias et des textes explicatifs) avec la possibilité d’interagir avec les contenus, par exemple un quiz.L’implémentation est un code libre et ouvert, facilement réutilisable par toute personne initiée à l’informatique.Cela constitue aussi en soi un outil d’apprentissage du code, en plus de l’acquisition de compétences en neuroscience intégrative, et c’est un levier de co-création.