Reproduce “Adaptive Retention & Correction: Test-Time Training for Continual Learning” on OpenResearch

About this paper

Haoran Chen, Micah Goldblum, Zuxuan Wu, Yu-Gang Jiang

ADAPTIVE RETENTION & CORRECTION (ARC) presents a test-time training approach for continual learning that addresses classifier bias towards recent tasks. The method significantly boosts model performance and reduces catastrophic forgetting by dynamically adapting the classification layer using identified past-task data during inference, without altering initial training procedures.

Key insights

• In large pretrained models, catastrophic forgetting primarily arises from classifier bias towards new tasks rather than degradation of the feature extractor, as representation layers retain knowledge well.
• Test-time data availability provides an opportunity to retrospectively leverage prior knowledge for bias correction, even without training-time memory buffers, addressing a gap in current memory-free continual learning.
• Dynamically identifying and using past-task samples during inference enables effective, memory-free mitigation of classifier bias without requiring changes to the original training setup.

Summary by alphaXiv.