Reproduce “Advancing Multiple Instance Learning with Continual Learning for Whole Slide Imaging” 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
Xianrui Li, Yufei Cui, Jun Li, Antoni B. Chan
Researchers at City University of Hong Kong, Huawei Canada, and Guangzhou Bingli Technology identified the attention network as the primary site of catastrophic forgetting in attention-based Multiple Instance Learning (MIL) for Whole Slide Imaging (WSI). They developed Attention Knowledge Distillation (AKD) and a Pseudo-Bag Memory Pool (PMP), which together improved Balanced Average Accuracy by up to 37% on large-scale WSI datasets compared to existing continual learning methods.
Key insights
- Decoupling experiments and gradient analysis revealed that the attention network, not the classifier, is the primary source of catastrophic forgetting in attention-based MIL models for WSI, characterized by its unbounded and rapidly updating gradients.
- A small, carefully selected subset of patches ('pseudo-bags') can effectively represent past knowledge for replay, making continual learning feasible for gigapixel WSIs without storing entire slides.
- Distilling knowledge directly from the attention layers, rather than just the final logits, is crucial for preserving diagnostic capabilities for old tasks in MIL models during continual learning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.