Reproduce “Overcoming Domain Drift in Online Continual Learning” on OpenResearch

About this paper

Fan Lyu, Daofeng Liu, Linglan Zhao, Zhang Zhang, Fanhua Shang, Fuyuan Hu, Wei Feng, Liang Wang

Researchers from the Chinese Academy of Sciences and Tencent's Youtu Lab developed Drift-Reducing Rehearsal (DRR), a method that addresses continual domain drift in Online Continual Learning. DRR combines centroid-based memory selection, a cross-task contrastive margin loss, and centroid distillation loss, achieving state-of-the-art performance across multiple benchmarks and visually demonstrating reduced feature space drift.

Key insights

• Continual domain drift, characterized by the gradual displacement of decision boundaries in feature spaces, is a primary cause of catastrophic forgetting in Online Continual Learning.
• Explicitly managing memory buffer content through centroid-based selection and regularizing feature representations via contrastive learning and distillation effectively mitigates domain drift.
• Maintaining stable and separated feature distributions for previously learned tasks is crucial for long-term knowledge retention in streaming data environments.

Summary by alphaXiv.