Meta Research

Recent advances in deep reinforcement learning require a large amount of data and result in representations that are often over specialized to the target task. In this work, we study the underlying potential causes for this specialization by measuring the similarity between representations trained on related, but distinct tasks. We use the recently proposed projection weighted Canonical Correlation Analysis (PWCCA) to examine the task dependence of visual representations learned across different embodied navigation tasks. Surprisingly, we find that slight differences in task have no measurable effect on the visual representation. We then empirically demonstrate that visual representations learned on one task can be effectively transferred to a different task. Finally, we show that if the tasks constrain the agent to spatially disjoint parts of the environment, differences in representation emerge, providing insight on how to design tasks that induce general, task-agnostic representations.