Recording
https://www.bilibili.com/video/BV1sh4y1A7eu/
Speaker
Stefano Palminteri
Bio
I am a Research Director (equivalent of Full Professor) and head of the Human Reinforcement Learning team, which is part of the Laboratoire de Neurosciences Cognitives et Computationelles. My goal is understanding how humans learn to make decisions at the behavioral, computational and neural levels. I am mainly (but not only!) interested in situations when decisions are based on past experience (a.k.a. reinforcement learning). In the last few years I mainly worked on two computational hypotheses, “relative value” and “learning bias”, concerning human reinforcement learning. In addition to extending these two frameworks, new lines of research in my team investigate social learning , the experience/description gap and, more recently, the intersection between cognitive science and artificial intelligence. I also enjoy questioning the epistemological and methodological foundations of decision-making, neuroeconomics and cognitive science research.
Abstract
In the present talk I will review research performed by my and other laboratories that highlight the existence of reinforcement leaning biases. I will first start by briefly introducing the reinforcement learning framework and propose a taxonomy of biases within the framework itself. Then I will review direct empirical evidence supporting the existence of two learning biases: positivity bias and value-normalization at both the behavioral, neural and clinical level. Finally, I will present unpublished results of simulation studies that try to answer the question: what are these reinforcement learning biases good for?