December 15, 2020

12:00 pm / 1:15 pm

Venue

https://wse.zoom.us/j/95370153750?pwd=blhib2R6YVFCbCsrWGhGcGFrVlJ4dz09

Ida Momennejad, PhD
Senior Reinforcement Learning     
at Microsoft Research NYC

Recorded Seminar:
https://wse.zoom.us/rec/share/JWQBlx7G4tLgDvr3UrUGCrZQTT9eIdPfJp3QzzJZeIlBFFXDCJPps1oWD78RrjUo.BTCzn9KBSLnr4Z8s?startTime=1608051066000    

Zoom Meeting link:
https://wse.zoom.us/j/95370153750?pwd=blhib2R6YVFCbCsrWGhGcGFrVlJ4dz09   

Title: Multi-scale Predictive Representations

Memory and planning rely onlearning the relational structure of experience. A century after ?latentlearning’ experiments summarized by Tolman, the larger puzzle of cognitive mapsremains elusive: how does the brain learn compact representations of relationalstructures to guide flexible behavior? I use reinforcement learning (RL) tostudy how humans learn predictive representations in memory and planning. Ishow behavioral, fMRI, and electrophysiology evidence that hippocampal andprefrontal hierarchies learn multi-scale predictive representations updated viaoffline replay. This approach advances the century old notion of cognitive mapsand can inform biologicallyinspired artificial agents as well as computationalpsychiatry.

BIO

Ida Momennejad is a senior researcher in reinforcementlearning at Microsoft Research NYC. Ida builds and tests neurally plausiblealgorithms for learning the structure of the environment such that it servesmemory,exploration, & planning. Her approach combines reinforcementlearningwith behavioral experiments, fMRI, & electrophysiology. Ida gother BSc in software engineering (Tehran, Iran), MSc in Philosophy of Science(Utrecht, Netherlands), PhD in psychology/computational neuroscience (Berlin,Germany). She was a postdoc at Princeton and an associate research scientist atColumbia BME.