Colloquium, Ida Momennejad, Princeton University

Date
Wed November 1st 2017, 3:35 - 5:00pm

“Predictive Representations in Memory and Planning” 

 

Abstract

As we navigate the world, what kinds of representations do the brain store in memory or retrieve to make decisions? How does the brain update these representations? I will present computational, behavioral, and fMRI evidence that memory and planning interact: (1) The brain builds and stores predictive maps (known as “successor representation”) in memory that cache multi-step dependencies at multiple scales, and uses them for fast and flexible planning; (2) Predictive representations and planning policies are updated via offline memory processes, such as replay; (3) Offline replay processes are triggered by the experience of uncertainty (unsigned prediction errors). This combination of predictive maps and uncertainty-triggered replay offers a powerful mechanism for segmenting and integrating the flow of experience into memory at different levels of abstraction. Such multi-level representations could serve as predictive maps that enable perception and planning at multiple time-scales.

Contact Phone Number