A multi-view consistency framework with semi-supervised domain adaptation

2024-10-01

Engineering applications of artificial intelligence

AI Summary

This work addresses the challenges of class ambiguity and prediction bias in semi-supervised domain adaptation, where labeled samples in the target domain are scarce. To mitigate these issues, the authors propose a multi-view consistency learning framework that leverages strong data augmentation to construct dual training views. In one view, a class-debiasing strategy refines the prediction distribution, while the other generates pseudo-negative labels to enhance inter-class separability. Additionally, cross-domain affinity learning is employed to align features of the same class across domains. By innovatively integrating debiased prediction, pseudo-negative labeling, and cross-domain alignment, the method significantly improves model generalization and class discriminability under limited annotation. Experiments on DomainNet and Office-Home demonstrate superior performance over existing approaches, effectively reducing labeling costs while enhancing cross-domain adaptation.

Technology Category

• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Computer Vision: Adversarial Attacks & Robustness
• Application Domains: Humanities & Computational Social Science

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling

Problem

• Semi-Supervised Domain Adaptation
• class bias
• limited labeled data
• feature space similarity
• domain adaptation

Innovation

• Multi-View Consistency
• Semi-Supervised Domain Adaptation
• Debiasing Strategy
• Pseudo-Negative Labels
• Cross-Domain Affinity Learning