Ida Momennejad, PhD
Title: Multi-scale Predictive Representations in Memory and Planning
Date: Wednesday, December 19th, 2018
Location: Herklotz Conference Center, Center for Neurobiology of Learning and Memory
How does the brain build models of the world in order to predict and plan the future? Tolman proposed cognitive maps that represent merely relationships between adjacent events in the environment. When faced with a decision, the one-step cognitive map would be used to simulate various paths to reward ‘online’, step-by-step. However, this proposal is computationally costly and inefficient for realistically large decision trees. An alternative proposal is that an efficient model of the environment generalizes multi-step structural relationships among states or events. Such a model can be updated by ‘offline’ memory processes; enabling fast and flexible planning. I use reinforcement learning, behavioral experiments, fMRI, and electrophysiology in humans to study how the brain maps the world. I show evidence that multi-step predictive representations, abstracted at multiple scales and updated via replay, govern planning behavior. These multi-scale maps are supported by the hippocampus and the prefrontal cortex, which also support non-spatial prospective memory and predictive representations. I will briefly discuss ongoing work in human electrophysiology and computational psychiatry. Taken together, this work suggests that multi-scale predictive representations may underlie memory, navigation, and planning.