文献类型：期刊文献

英文题名：Confusion distillation for Continual Self-Supervised Learning

作者：Liu, Ganghao[1,2];Zhou, Qin[1,2];Fu, Zhiling[1,2];Wang, Zhe[1,2]

机构：[1]Minist Educ, Key Lab Smart Mfg Energy Chem Proc, Shanghai 200237, Peoples R China;[2]East China Univ Sci & Technol, Dept Comp Sci & Engn, Shanghai 200237, Peoples R China

年份：2026

卷号：180

外文期刊名：PATTERN RECOGNITION

收录：;EI(收录号：20262420878221);WOS:【SCI-EXPANDED(收录号:WOS:001794420300001)】；

基金：This work is supported by National Natural Science Foundation of China under Grant Nos. 62476087 and 62201341, Shanghai Municipal Education Commission's Initiative on Artificial Intelligence-Driven Reform of Scientific Research Paradigms and Empowerment of Discipline Leapfrogging.

语种：英文

外文关键词：Self-supervised learning; Continual learning; Knowledge distillation; Task confusion; Self-supervised learning; Continual learning; Knowledge distillation; Task confusion

摘要：Self-Supervised Learning (SSL) has achieved competitive performance across a range of computer vision tasks, but it typically relies on the assumption that training data are independent and identically distributed. In real-world scenarios, data often arrive in a non-stationary and streaming manner, posing significant challenges to standard SSL approaches. Continual Self-Supervised Learning (CSSL) offers a promising solution by enabling models to learn from unlabeled data streams. However, existing methods often suffer from limited adaptability due to over-constraining regularization, and from semantic interference between tasks, leading to representational entanglement known as task confusion. To mitigate these limitations, we propose a self-supervised confusion distillation framework for unsupervised continual learning. The framework introduces an auxiliary self-supervised branch to extract task-adaptive representations, where its decoupled design from the backbone mitigates representational interference across sequential tasks. To combat catastrophic forgetting and integrate knowledge across tasks, we introduce a confusion distillation mechanism that fuses historical and current representations via feature concatenation and self-distillation. Furthermore, to prevent task confusion and ensure model stability, we devise a CutMix-based augmentation strategy with a replay mechanism. This strategy ensures that the memory buffer captures diverse task-specific characteristics, facilitating effective cross-task mixing and mitigating task confusion throughout the incremental learning process. Extensive experiments on CIFAR100 and ImageNet-Sub benchmarks demonstrate that our method consistently outperforms existing CSSL approaches across multiple SSL paradigms and task configurations, highlighting its effectiveness and generalizability.