Wed March 22nd 2023, 3:45 - 5:00pm
Stanford Neurosciences Building, John A. and Cynthia Fry Gunn Rotunda, E241
290 Jane Stanford Way

Dr. Flora Bouchacourt

Dr. Flora Bouchacourt, Postdoc, Carney Institute for Brain Science, Brown University

Title: The flexible brain: working memory and task learning in the prefrontal cortex

Abstract: One particular promise of computational psychiatry is to build a theory of brain function disentangling the etiology of mental disorders. For example, disruption of working memory - our ability to hold and use information in mind - and impairments in adapting to new situations (“tasks”), are known markers of prefrontal dysfunctions and schizophrenia. There is currently no satisfying treatment to restore these cognitive functions. 

Many daily and habitual tasks broadly involve learning, i.e. adjusting ongoing behavior in light of experiences within a known task. One way to conceptualize the ability to adapt to a new task is as a process of “meta-learning”, i.e. producing a strategy tailored for learning in the task. It is thought to be solved by the brain by building fast internal representations for task states in the prefrontal cortex (“task state spaces”), yet with largely unknown mechanisms. Bridging the gap between behavior and neurobiological constraints requires unifying models and a multidisciplinary endeavor I wish to be part of.

I will first show how we can model the trade-off between the flexibility and the limited capacity of working memory. Second, I will present how we can model humans and monkeys’ ability to learn and switch between rules for a known task, and what we learn from it. A goal for my future lab is to show that these cognitive functions ultimately go together to allow meta-learning of new tasks. One theoretical insight reframes meta-learning as a well-known problem called a Partially Observable Markov Decision Process, which addresses the general problem of choice in tasks with unknown states. This framework allows for a combination of multilevel tools, from Bayesian reinforcement learning to detailed neural network modeling. The goal for the first years of my lab is to build on this insight to develop a theory of fast meta-learning in the brain, and to test it using neuroimaging. In collaboration with psychiatrists, I also plan to examine the mechanisms of the related cognitive impairments in psychosis.