Dynamic Cortico-Hippocampal Interactions for Flexible Goal-Driven Agents
Abstract
How does the human brain enable us to flexibly adapt to novel situations, applying hard-won prior experience in ways that apply to new challenges? Current AI systems continue to fail at capturing this central feature of human intelligence. We propose to leverage our extensive understanding of the neural organization of the human brain to suggest novel solutions to thisperennial challenge, using a synergistic combination of biologically-based computational models and cutting-edge neuroimaging techniques, including intracranial recordings and fMRI. Our approach involves four central hypotheses: 1) The brain learns and processes structure separate from content , along separate neural pathways. This enables abstract structural representations tobe flexibly recombined with novel content, while transferring structural inferences and implications to these novel domains. 2) Predictive learning can learn an abstract structural forward model of how events unfold over time (i.e., a situation model ), providing a biologically plausible mechanism for how the neural mechanisms that enable flexible behavior are generated. 3) The ability to direct behavior toward desired outcomes and goals requires the ability to reversethis model and figure out the actions required to achieve those goals. The hippocampus can help solve this problem by encoding the overarching situation model that led to a given outcome, and then later recalling that situation model for use as a plan to guide behavior toward that outcome, when it is desired again. This suggests a novel type of strategic control over hippocampalencoding and recall, and a critical role for episodic memory in enabling flexible human behavior. 4) The prefrontal cortex can robustly maintain abstract plans to guide behavior, based on representations learned through longer-timescale predictive learning. Each of these hypotheses has existing empirical and computational support, but this project is the first to put all of these elements together in a way that can finally make significant progress toward understandingflexible goal-driven human behavior.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 11, 2020
- Source ID
- N000142012578
Entities
People
- Randall O Reilly
Organizations
- Office of Naval Research
- United States Navy
- University of California, Davis