Active Predictive Coding Models of Cortical and Subcortical Function
Abstract
One of the central problems in neuroscience today is how the neocortex, with its laminar architecture, its reciprocal connections between areas and its interactions with the thalamus, hippocampus and other subcortical regions, performs functions as diverse as visual perception, motor planning, episodic recall, and abstract reasoning. We propose to investigate the hypothesis that the neocortex implements active predictive coding (APC), a new model for perception and action that postulates that each cortical area estimates both sensory states and actions (including potentially abstract actions internal to the cortex). The cortex as a whole learns to predict the sensory consequences of actions at multiple hierarchical levels. We will explore, using simulations and comparisons to published experimental data, the following questions regarding the APC hypothesis- (1) What is the role of top-down feedback connections from higher to lower cortical areas in prediction, active perception, hierarchical planning, and compositional learning. (2) What are the roles of cortical deep versus superficial layer neurons and thalamic neurons in computing predictions and conveying prediction errors for inference and learning. (3) How can uncertainty be represented in the APC model using precision-based or sampling-based representations. (4) How is hierarchical action selection implemented via cortical interactions with the basal ganglia, thalamus and other subcortical structures. (5) How do the interactions between the cortex and hippocampus allow binding, episodic memories, and fast transfer learning. We expect our research to result in a new general framework for artificial intelligence that emphasizes learning hierarchical sensory-motor models, with AI and robotics applications such as efficient perception, learning, planning and closed-loop control that are central to the missions of the Air Force and DoD.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 06, 2025
- Source ID
- FA95502410313
Entities
People
- Rajesh Rao
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Washington