GLC++: Source-Free Universal Domain Adaptation Through Global-Local Clustering and Contrastive Affinity Learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural networks struggle with covariate and category shifts, impacting performance.
• Source-Free Domain Adaptation (SFDA) offers solutions but is often limited to closed-set scenarios.
• Existing methods fail to effectively distinguish between known and unknown data categories in open-set scenarios.

Purpose of the Study:

• To explore Source-Free Universal Domain Adaptation (SF-UniDA) for classifying known and unknown data under category shifts.
• To propose novel clustering techniques to improve model robustness and accuracy in domain adaptation.
• To enhance the identification and clustering of distinct unknown categories.

Main Methods:

• Developed Global and Local Clustering (GLC) with adaptive global clustering and local k-NN for mitigating negative transfer.
• Introduced GLC++, an evolution of GLC, incorporating contrastive affinity learning for improved unknown category identification.
• Evaluated GLC and GLC++ on multiple benchmarks across various category shift scenarios.

Main Results:

• GLC and GLC++ demonstrated superior performance in challenging open-partial-set scenarios, outperforming existing methods like GATE.
• GLC++ significantly improved novel category clustering accuracy in open-set scenarios compared to GLC.
• The integrated contrastive learning strategy boosted the performance of both GLC and other existing domain adaptation methodologies.

Conclusions:

• SF-UniDA, particularly with GLC and GLC++, offers a robust solution for deep learning models facing domain and category shifts.
• The proposed methods effectively handle both known and unknown data categories, improving overall model adaptability.
• Contrastive learning integration presents a promising direction for advancing source-free domain adaptation techniques.