Les séminaires du lundi accueillent Valentin WYART, Directeur de Recherche INSERM au Laboratoire de Neurosciences Cognitives et Computationnelles (LNC2) et Professeur attaché en Intelligence Artificielle à l'Ecole Normale Supérieure - Université PSL à Paris.

Titre : Bug or feature? How computation noise shapes learning and decision-making under uncertainty in humans and artificial neural networks

Abstract : Making sense of uncertain and volatile environments, a cognitive process modeled across domains in terms of statistical inferences, constitutes a difficult yet ubiquitous challenge for human intelligence. Beside sensory errors and exploratory choices, recent research has identified the limited precision of inferences themselves as a key contribution to the variability and suboptimality of perceptual and reward-guided decisions made under uncertainty. In this talk, I will present and discuss the theoretical and experimental evidence obtained by my group which, taken together, provides key insights into the origin, impact and function of this 'computation noise' for human learning and decision-making. First, I will show that computation noise acts as a domain-general constraint that shapes how humans solve decision problems under uncertainty, and explains a range of strategies and biases otherwise seen as idiosyncratic. Second, moving beyond the classical description of internal noise as a constraint on decision accuracy, I will present new results that delineate not only the costs, but also the emergent benefits of computation noise for artificial neural networks tested in a range of uncertain conditions.

RDV Espace POUILLON - Accueil café 15 minute avant.