Reproduce “FOSTER: Feature Boosting and Compression for Class-Incremental Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, De-Chuan Zhan
Researchers at Nanjing University developed FOSTER, a two-stage class-incremental learning paradigm employing feature boosting and compression to address catastrophic forgetting and parameter growth. The method achieved state-of-the-art performance, including up to 3.11% higher average incremental accuracy on CIFAR-100 compared to prior methods, while maintaining model efficiency.
Key insights
- Applying gradient boosting principles enables new modules to specifically address discrepancies (residuals) from prior knowledge, leading to refined feature learning across all categories.
- Calibrating logits and enhancing features for both old and new classes during the boosting phase is critical for mitigating classification bias and fostering comprehensive knowledge acquisition.
- A balanced knowledge distillation approach can effectively compress dynamic, expanded model architectures into a single, efficient backbone without significant performance loss, addressing parameter growth.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.