Reproduce “Multi-Granularity Class Prototype Topology Distillation for Class-Incremental Source-Free Unsupervised Domain Adaptation” on OpenResearch

About this paper

Peihua Deng, Jiehua Zhang, Xichun Sheng, Chenggang Yan, Yaoqi Sun, Ying Fu, Liang Li

Researchers introduce Multi-Granularity Class Prototype Topology Distillation (GROTO), an algorithm addressing Class-Incremental Source-Free Unsupervised Domain Adaptation (CI-SFUDA) by mitigating negative transfer and catastrophic forgetting. GROTO achieves state-of-the-art performance on three benchmark datasets, improving average final accuracy by up to 5.2% on Office-31-CI and 1.3% on Office-Home-CI compared to prior methods.

Key insights

• Robustly identifying positive classes through hybrid knowledge accumulation (source similarity and target probability) effectively reduces negative transfer from irrelevant source classes in CI-SFUDA.
• Utilizing a combination of coarse-grained (stable source/target representatives) and fine-grained (highly confident target features) prototypes leads to more reliable pseudo-labels for target representation learning.
• Distilling the topological relationships of class prototypes between the comprehensive source model and the incrementally learned target model helps maintain old knowledge and prevents decision boundaries from overfitting to new classes, mitigating catastrophic forgetting.

Summary by alphaXiv.