Adversarial Projections to Tackle Support-Query Shifts in Few-Shot Meta-Learning

Aroof Aimen, Bharat Ladrecha, Narayanan Chatapuram Krishnan

Abstract: Popular few-shot Meta-learning (ML) methods presume that a task's support and query data are drawn from a common distribution. A recent work relaxed this assumption to propose a few-shot setting where the support and query distributions differ, with disjoint yet related meta-train and meta-test support-query shifts (SQS). We relax this assumption further to a more pragmatic SQS setting (SQS+) where the meta-test SQS is unknown and need not be related to the meta-train SQS. The state-of-the-art solution to address SQS is transductive, requiring unlabelled meta-test query data to bridge the support and query distribution gap. In contrast, we propose a theoretically grounded inductive solution - Adversarial Query Projection (AQP) for addressing SQS+ and SQS. AQP can be easily integrated into the popular ML frameworks. Exhaustive empirical investigations on benchmark datasets and their extensions, different ML approaches, and architectures establish AQP's efficacy in handling SQS+ and SQS.

Keywords: Meta-learning, Task, Support, Query, Projection, Shift, Adversarial

One-sentence Summary: We propose an Inductive Solution to Address Support Query Shifts in Meta-learning.

Reproducibility Checklist: Yes

Broader Impact Statement: Yes

Paper Availability And License: Yes

Code Of Conduct: Yes

Reviewers: Aroof Aimen,[e-mail]

Code And Dataset Supplement: https://github.com/Few-Shot-SQS/adversarial-query-projection