Our new X account is live! Follow @wizwand_team for updates
WorkDL logo mark

Domain Adaptation with Asymmetrically-Relaxed Distribution Alignment

About

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.

Yifan Wu, Ezra Winston, Divyansh Kaushik, Zachary Lipton• 2019

Related benchmarks

TaskDatasetResultRank
Unsupervised Domain AdaptationOffice-Home RS-UT label shifts
Accuracy (Rw -> Pr)68.56
16
Domain AdaptationOfficeHome RS->UT
Accuracy (Rw -> Pr)68.56
14
Behavior DecodingNHP motor cortex (Cross-Day)
R^2-0.1736
9
Behavior DecodingNHP motor cortex (Inter-Subject)
R^2-19.59
9
Neural Decodingrat hippocampus dataset (across sessions)
R^2-0.0288
7
Showing 5 of 5 rows

Other info

Follow for update