Insights on Visual Representations for Embodied Navigation Tasks

Recent advances in deep reinforcement learning require a large amount of training data and generally 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 for both SqueezeNet and ResNet architectures. We then empirically demonstrate that visual representations learned on one task can be effectively transferred to a different task. Interestingly, we show that if the tasks constrain the agent to spatially disjoint parts of the environment, differences in representation emerge for SqueezeNet models but less-so for ResNets, suggesting that ResNets feature inductive biases which encourage more task-agnostic representations, even in the context of spatially separated tasks. We generalize our analysis to examine permutations of an environment and find, surprisingly, permutations of an environment also do not influence the visual representation. Our analysis provides insight on the overfitting of representations in RL and provides suggestions of how to design tasks that induce task-agnostic representations.