Affiliation: Laboratoire de Neurosciences Cognitives et Computationnelles, Institut National de la Santé et de la Recherche Médicale (INSERM), Paris, France.

Abstract: Since its inception in 2017, one of our main research goals has been to propose, computationally formalize, and behaviorally validate the existence of biases in the way humans learn from rewards and punishments. In particular, we have identified two key processes: relative outcome encoding and a positivity/confirmation bias in learning. We have tested these ideas across a wide range of experimental designs, which has helped refine our computational understanding of these biases. Their empirical validation has systematically relied on what we consider methodological gold standards in computational cognitive modeling (coupling parsimony with falsification). However, fueled by strict open data practices and the broad interest generated by these questions, many laboratories have sought to challenge our conclusions over the years, sometimes proposing radically different computational interpretations of the behavioral effects that are otherwise widely replicated. In this talk, I will recount the story of these challenges and our responses. I will argue that, taken together, this process provides unusually strong evidence that the reinforcement learning biases we have identified withstand a level of scrutiny that is both rare and highly valuable for the field.

Webex-Link: