Dynamic Cortico-Hippocampal Interactions for Flexible Goal-Driven Agents
Abstract
How does the human brain enable us to generate and execute a plan across long timescales, applying hard-won prior experience in ways that apply to new challenges? State of the art artificial intelligence systems require massive amounts of training data and still struggle to emulate this central feature of human intelligence. Neuroscience can potentially suggest novel solutions to this perennial challenge, but current thinking in neuroscience focuses on single brain regions like the hippocampus and we currently lack the theoretical tools and empirical data to understand how brain regions interact across long timescales to generate intelligent, goal-driven behavior. We propose a synergistic program of computational modeling and cutting-edge neuroscience approaches to bridge this gap. Central to our #Content, Goals, and Structure# (CoGS) framework is the idea that intelligent behavior depends on strategically timed interactions between neocortical systems that gradually learn about events, environments, and tasks and a fast-learning hippocampal system optimized for episodic memory. We will test and refine this framework with three lines of computational modeling research,as well as human neuroscience research using neuroimaging and intracranial single-unit and local field potential recordings. Our computational research will focus on: (1) development of our biologically-based models of cortico-hippocampal interactions during learning and navigation of graph structures and cognitive maps, (2) development of more abstract computational tools that utilize widelyavailable machine learning frameworks to bridge the gap between neuroscience and conventional AI, and (3) #stimulus computable# models that learn environmental structure from sensors and autonomously navigate. Our empirical research will address: (1) Mechanisms and precise timing of cortico-hippocampal interactions using intracranial recordings, (2) Determinants and consequences of cortico-hippocampal interactions during spatial learning using virtual reality and functional magnetic resonance imaging. The synergistic application of cutting-edge human neuroscience methods and computational modeling at multiple levels of abstraction will lead to significant breakthroughs in state-of-the-art AI, resulting in transformative advances in military systems for analysis of data streams andhuman machine teaming.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 15, 2024
- Source ID
- N000142412325
Entities
People
- Charan Ranganath
Organizations
- Office of Naval Research
- United States Navy
- University of California, Davis