Domain Adaptation with Asymmetrically-Relaxed Distribution Alignment

TL;DR

This paper introduces an asymmetrically-relaxed distribution alignment method for domain adaptation, addressing limitations of existing adversarial approaches by providing theoretical guarantees and demonstrating empirical improvements on synthetic and real datasets.

Contribution

It proposes a novel asymmetrically-relaxed alignment technique that overcomes key limitations of standard adversarial domain adaptation methods, with theoretical analysis and empirical validation.

Findings

The new method outperforms standard approaches on synthetic datasets.

Empirical benefits observed on real-world datasets.

Theoretical analysis clarifies the assumptions for principled adaptation.

Abstract

Domain adaptation addresses the common problem when the target distribution generating our test data drifts from the source (training) distribution. While absent assumptions, domain adaptation is impossible, strict conditions, e.g. covariate or label shift, enable principled algorithms. Recently-proposed domain-adversarial approaches consist of aligning source and target encodings, often motivating this approach as minimizing two (of three) terms in a theoretical bound on target error. Unfortunately, this minimization can cause arbitrary increases in the third term, e.g. they can break down under shifting label distributions. We propose asymmetrically-relaxed distribution alignment, a new approach that overcomes some limitations of standard domain-adversarial algorithms. Moreover, we characterize precise assumptions under which our algorithm is theoretically principled and demonstrate empirical benefits on both synthetic and real datasets.