Analyzing Visual Representations in Embodied Navigation Tasks

TL;DR

This paper investigates how visual representations in embodied navigation tasks are affected by task differences, revealing that slight task variations do not alter representations and that transfer learning is effective across tasks.

Contribution

It introduces a methodology using PWCCA to analyze the task dependence of visual representations in embodied navigation and demonstrates transferability across tasks.

Findings

Slight task differences do not significantly change visual representations.

Visual representations can be transferred effectively between related tasks.

PWCCA is useful for analyzing representation similarity in embodied navigation.

Abstract

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 present a methodology to study the underlying potential causes for this specialization. We use the recently proposed projection weighted Canonical Correlation Analysis (PWCCA) to measure the similarity of visual representations learned in the same environment by performing different tasks. We then leverage our proposed methodology to examine the task dependence of visual representations learned on related but distinct 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.