Implementation of a Model of the Cortex Basal Ganglia Loop

TL;DR

The paper presents a minimal, modular implementation of the cortico-basal ganglia-thalamo-cortical loop as a reusable component for action selection in brain-inspired architectures. It tests a hypothesis where the cortex predicts actions while the basal ganglia learn a Go/NoGo policy to regulate execution, with a simplified thalamus gating mechanism. The implementation includes a cortex with an output predictor, moderator, and selector, and a basal ganglia reinforcement learner, evaluated on a delayed reward task and compared to a standard DQN baseline. Results show the BG can learn Go/NoGo decisions and the cortex can improve action prediction over time, offering a reusable module for larger cognitive architectures, while acknowledging biological plausibility limits and outlining avenues for future extension to more complex cognitive tasks.

Abstract

This article presents a simple model of the cortex-basal ganglia-thalamus loop, which is thought to serve for action selection and executions, and reports the results of its implementation. The model is based on the hypothesis that the cerebral cortex predicts actions, while the basal ganglia use reinforcement learning to decide whether to perform the actions predicted by the cortex. The implementation is intended to be used as a component of models of the brain consisting of cortical regions or brain-inspired cognitive architectures.

Figures (3)

• Figure 1: Overall Architecture
• Figure 2: Learning Results Horizontal axis: the number of episodes (×100) Red line: average reward when Go is selected Blue line: average reward Yellow line: mean squared error of the cortical prediction
• Figure 3: RL Results Horizontal axis: the number of episodes (×100) Vertical axis: average reward