University of Waterloo

Purpose

The aim of this project is to take an existing high-level abstract model of how reinforcement learning (RL) occurs, and ground it in detailed biology and molecular-level dynamics of the basal ganglia. That is, the project team will show how these low-level molecular details combine with biological structures (neurons and synapses in the basal ganglia) to produce overt behavior. Animals and humans can select actions that are continuous, e.g. how hard to turn a knob, or that are extended through time, e.g. reach and grasp trajectories. The basal ganglia in the mammalian brain is thought to be associated with this kind of action selection, acting as a control mechanism for initiating actions. Models of the basal ganglia, however, are generally based on reinforcement learning theory that treat both action spaces and time as inherently discrete. However, in recent work (funded under the NRC’s AI for Logistics program), a completely continuous theory has been developed. This theory, while compatible with the biology of the basal ganglia, is specified at a high level with abstract components and variables. This project proposes to connect this theory to the actual biological details of the basal ganglia at the level of biological connectivity, neurotransmitters, and other molecules. The primary focus will be on two areas: the biological realism of the inputs to the basal ganglia, including reward representations; and the biological realism of the continuous outputs from the basal ganglia, enabling selection in continuous action spaces. Given the resulting model, the team will then explore the effects of varying low-level molecular details (such as the concentrations of GABA, glutamate, cholesterol, etc.) on high-level behaviour such as navigating T-mazes and the Morris Water Maze.