Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Zero-shot domain adaptation based on dual-level mix and contrast

Yu Zhe, Jun Sakuma

TL;DR

This paper tackles zero-shot domain adaptation where the target domain lacks labeled samples for the task of interest. It introduces Dual Mixup Contrastive Learning (DMCL), which combines dual mixup to generate intermediate samples across domains and tasks, extended domain adversarial training, and dual-level contrastive learning to disentangle domain and task information. Empirical results on X-NIST and Office-Home show competitive or state-of-the-art performance with ablations confirming the necessity of each component, while visualizations indicate substantially domain-invariant features and clear separation of ToI versus IrT. DMCL offers a scalable approach to reducing task bias and achieving robust transfer under domain shifts without relying on heavy generative models.

Abstract

Zero-shot domain adaptation (ZSDA) is a domain adaptation problem in the situation that labeled samples for a target task (task of interest) are only available from the source domain at training time, but for a task different from the task of interest (irrelevant task), labeled samples are available from both source and target domains. In this situation, classical domain adaptation techniques can only learn domain-invariant features in the irrelevant task. However, due to the difference in sample distribution between the two tasks, domain-invariant features learned in the irrelevant task are biased and not necessarily domain-invariant in the task of interest. To solve this problem, this paper proposes a new ZSDA method to learn domain-invariant features with low task bias. To this end, we propose (1) data augmentation with dual-level mixups in both task and domain to fill the absence of target task-of-interest data, (2) an extension of domain adversarial learning to learn domain-invariant features with less task bias, and (3) a new dual-level contrastive learning method that enhances domain-invariance and less task biasedness of features. Experimental results show that our proposal achieves good performance on several benchmarks.

Zero-shot domain adaptation based on dual-level mix and contrast

TL;DR

This paper tackles zero-shot domain adaptation where the target domain lacks labeled samples for the task of interest. It introduces Dual Mixup Contrastive Learning (DMCL), which combines dual mixup to generate intermediate samples across domains and tasks, extended domain adversarial training, and dual-level contrastive learning to disentangle domain and task information. Empirical results on X-NIST and Office-Home show competitive or state-of-the-art performance with ablations confirming the necessity of each component, while visualizations indicate substantially domain-invariant features and clear separation of ToI versus IrT. DMCL offers a scalable approach to reducing task bias and achieving robust transfer under domain shifts without relying on heavy generative models.

Abstract

Zero-shot domain adaptation (ZSDA) is a domain adaptation problem in the situation that labeled samples for a target task (task of interest) are only available from the source domain at training time, but for a task different from the task of interest (irrelevant task), labeled samples are available from both source and target domains. In this situation, classical domain adaptation techniques can only learn domain-invariant features in the irrelevant task. However, due to the difference in sample distribution between the two tasks, domain-invariant features learned in the irrelevant task are biased and not necessarily domain-invariant in the task of interest. To solve this problem, this paper proposes a new ZSDA method to learn domain-invariant features with low task bias. To this end, we propose (1) data augmentation with dual-level mixups in both task and domain to fill the absence of target task-of-interest data, (2) an extension of domain adversarial learning to learn domain-invariant features with less task bias, and (3) a new dual-level contrastive learning method that enhances domain-invariance and less task biasedness of features. Experimental results show that our proposal achieves good performance on several benchmarks.
Paper Structure (17 sections, 15 equations, 4 figures, 3 tables)

This paper contains 17 sections, 15 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Contribution
  4. Preliminary
  5. Problem setting
  6. Domain adversarial training
  7. Mixup
  8. Methodology
  9. Overview
  10. Dual Mixup for intermediate samples
  11. Domain adversarial training with dual mixup
  12. Dual contrastive learning for disentanglement
  13. Experiments
  14. Experimental results on two benchmarks
  15. Visualizations
  16. ...and 2 more sections

Figures (4)

  • Figure 1: Zero-shot domain adaptation
  • Figure 2: An illustration of our proposal DMCL. Here GRL refers to a gradient reversal layer. The red arrow indicates the adversarial learning procedure, the green arrow indicates a standard training procedure, and the pink arrow indicates the contrastive learning objectives.
  • Figure 3: An illustration of dual contrastive learning, here the dashed line indicates apply mixup on two samples.
  • Figure 4: Visualization results when domain shift is domain G to domain N. At the top, the ToI is MNIST, IrT is Fashion-MNIST. At the bottom, the ToI is MNIST, IrT is EMNIST.