Building expertise through task-specific representational alignment in biological and artificial neural networks

Ashug Gurijala, Michael Nguyen Jr, John O'Doherty, Sneha Aenugu

Track: Extended Abstract Track

Keywords: Representational Learning, Reinforcement Learning, Cognitive Science, Neural Networks, Skill Acquisition

Abstract: Humans can generate both rapid and accurate responses in diverse tasks by building perceptuo-motor expertise through practice. Expert responses are robust to task-irrelevant distractors and state-space nuisance. In this paper, we investigate the representational transformations that guide skill acquisition in both humans and artificial agents. Specifically, we investigate the hypothesis that the evolution of task-specific efficient representational coding emerges in the higher layers of the visuo-motor hierarchy in biological and artificial networks. Towards this end, we built a custom shooter game with the specific aim of introducing maximal variance in perceptual state spaces, in which the development of expertise entails building robustness to such state-space distortions. Deep reinforcement learning agents playing the game develop representational alignment with the task-relevant features in higher layers late in the training process, with the lower layers remaining agnostic to the task. We aim to investigate parallel representational alignment in humans through longitudinal neural recordings to precisely probe the evolution of representational bottlenecks that result in the formation of expertise.

Submission Number: 122