Solving motion coherence with biologically plausible agents

Abstract

Artificial Neural Networks (ANNs) have been increasingly used in attempts to mechanis- tically explain the inner workings of human cognition. Motion coherence is a a well-known paradigm whose computation requires closing the Perception-Action cycle. Here, artificial agents with short-term memory and simulated vision learn how to solve the task through reinforcement learning. The agents are able to solve the task in a way comparable to human participants. The results are consistent with behavioral and psychological measures, and bear electrophysiological similarities. The results suggest that the proposed framework is a robust model choice for solving tasks that require closing the Perception-Action cycle.