Learning useful representations for shifting tasks and distributions

TL;DR

This paper argues that representations learned from multiple training episodes with different random seeds are richer and more adaptable to distribution shifts than those learned from a single episode, supported by theoretical insights and experiments.

Contribution

It introduces the idea that concatenating representations from multiple training runs enhances adaptability to new distributions, highlighting the limitations of single-episode learned representations.

Findings

Concatenated representations outperform single-run representations on new distributions.

Multiple training episodes produce more diverse and informative representations.

Single training episodes tend to produce less redundant and less adaptable features.

Abstract

Does the dominant approach to learn representations (as a side effect of optimizing an expected cost for a single training distribution) remain a good approach when we are dealing with multiple distributions? Our thesis is that such scenarios are better served by representations that are richer than those obtained with a single optimization episode. We support this thesis with simple theoretical arguments and with experiments utilizing an apparently na\"{\i}ve ensembling technique: concatenating the representations obtained from multiple training episodes using the same data, model, algorithm, and hyper-parameters, but different random seeds. These independently trained networks perform similarly. Yet, in a number of scenarios involving new distributions, the concatenated representation performs substantially better than an equivalently sized network trained with a single training run. This proves that the representations constructed by multiple training episodes are in fact different. Although their concatenation carries little additional information about the training task under the training distribution, it becomes substantially more informative when tasks or distributions change. Meanwhile, a single training episode is unlikely to yield such a redundant representation because the optimization process has no reason to accumulate features that do not incrementally improve the training performance.