Mirror Sample Based Distribution Alignment for Unsupervised Domain Adaption

Yin Zhao, Minquan Wang, Longjun Cai

Abstract:

Unsupervised Domain Adaption has great value in both machine learning theory and applications.
The core issue is how to minimize the domain shift. Motivated by the more and more sophisticated distribution alignment methods in sample level, we introduce a novel concept named (virtual) mirror, which represents the counterpart sample in the other domains.The newly-introduced mirror loss using the virtual mirrors establishes the connection cross domains and pushes the virtual mirror pairs together in the aligned representation space. Our proposed method does not align the samples cross domains coarsely or arbitrarily, thus does not distort the internal distribution of the underline distribution and brings better asymptotic performances.
Experiments on several benchmarks validate the superior performance of our methods.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Reviewed Version (pdf): https://openreview.net/references/pdf?id=5BsPZgH9JY