Zero-Shot Generalization of Vision-Based RL Without Data Augmentation

TL;DR

This paper introduces ALDA, a model that enables zero-shot generalization in vision-based RL without data augmentation by leveraging latent disentanglement and associative memory, inspired by neuroscience.

Contribution

The paper proposes ALDA, a novel approach combining latent disentanglement and associative memory to achieve zero-shot generalization in RL without data augmentation.

Findings

ALDA achieves zero-shot generalization on challenging task variations.

Data augmentation is shown to be a form of weak disentanglement.

The approach reduces computational and data collection costs.

Abstract

Generalizing vision-based reinforcement learning (RL) agents to novel environments remains a difficult and open challenge. Current trends are to collect large-scale datasets or use data augmentation techniques to prevent overfitting and improve downstream generalization. However, the computational and data collection costs increase exponentially with the number of task variations and can destabilize the already difficult task of training RL agents. In this work, we take inspiration from recent advances in computational neuroscience and propose a model, Associative Latent DisentAnglement (ALDA), that builds on standard off-policy RL towards zero-shot generalization. Specifically, we revisit the role of latent disentanglement in RL and show how combining it with a model of associative memory achieves zero-shot generalization on difficult task variations without relying on data augmentation. Finally, we formally show that data augmentation techniques are a form of weak disentanglement and discuss the implications of this insight.