Reproduce “Reusing the Task-specific Classifier as a Discriminator: Discriminator-free Adversarial Domain Adaptation” on OpenResearch

About this paper

Lin Chen, Huaian Chen, Zhixiang Wei, Xin Jin, Xiao Tan, Yi Jin, Enhong Chen

This paper introduces a Discriminator-free Adversarial Learning Network (DALN) that reuses the task-specific classifier as a discriminator for unsupervised domain adaptation. It proposes a novel Nuclear-norm Wasserstein Discrepancy (NWD) that inherently satisfies K-Lipschitz continuity, enabling robust domain alignment and category distinguishment and achieving state-of-the-art results on several benchmarks.

Key insights

• A classifier trained on the source domain inherently possesses an 'implicit discriminative ability' to distinguish between source-like and target-like predictions based on their certainty and correlation patterns.
• The proposed Nuclear-norm Wasserstein Discrepancy (NWD) serves as a theoretically sound critic function that naturally satisfies the K-Lipschitz constraint, eliminating the need for complex regularization techniques like weight clipping or gradient penalties.
• NWD unifies domain alignment and category distinguishment by encouraging both high prediction determinacy (for source) and diversity (for target), effectively combating mode collapse and improving feature transferability.

Summary by alphaXiv.