Student First Author: Yes
Previously Published: No
Abstract: Meta-learning is a paradigm whereby an agent is trained with the specific goal of fast adaptation. With promising recent results, it is an effective way to use prior knowledge for future learning. However, most of the prominent successes focus entirely on adapting to similar tasks, from a unimodal distribution. This drastically differs from the real world, where problems may be more diverse. In this paper we address this issue, and provide a simple approach to solve it, which we call Taming the Herd. The Herd refers to a population of agents, each specializing on a subset (or mode) of the task distribution. At test time, we automatically allocate the appropriate member of the Herd and thus perform comparably with an oracle trained solely on those tasks. We apply our approach to both MAML and PEARL, and demonstrate its efficacy on a simple yet challenging multi-modal task distribution.
TL;DR: A simple approach for meta-learning with a multi-modal task distribution.