Reproduce “Hierarchical Prompts for Rehearsal-free Continual Learning” on OpenResearch

About this paper

Yukun Zuo, Hantao Yao, Lu Yu, Liansheng Zhuang, Changsheng Xu

Hierarchical Prompts (H-Prompts) introduces a rehearsal-free continual learning method that leverages a frozen Vision Transformer backbone with structured learnable prompts to mitigate catastrophic forgetting. This approach achieves 87.8% average accuracy on Split CIFAR-100 and 70.6% on Split ImageNet-R, outperforming state-of-the-art exemplar-based and other prompt-based methods without storing historical data.

Key insights

• Explicitly structuring prompts into hierarchical categories (class, task, general) effectively combats catastrophic forgetting by enabling distinct roles for knowledge preservation and acquisition.
• Modeling class distributions as learnable 'class prompts' through adversarial training allows for rehearsal-free replay of past knowledge, bypassing the need for exemplar storage.
• Combining class-level distribution modeling, cross-task knowledge integration, and self-supervised generalized knowledge learning provides a robust framework for continual adaptation.

Summary by alphaXiv.