Domain Adaptation with Asymmetrically-Relaxed Distribution Alignment
This work addresses domain adaptation for machine learning applications where data distributions shift, but it appears incremental as it builds on existing adversarial approaches with specific theoretical refinements.
The paper tackles the problem of domain adaptation where target test data distribution drifts from the source training distribution, proposing asymmetrically-relaxed distribution alignment to overcome limitations of standard domain-adversarial algorithms, and demonstrates empirical benefits on synthetic and real datasets.
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.