Reproduce “OpenIncrement: A Unified Framework for Open Set Recognition and Deep 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

Jiawen Xu, Claas Grohnfeldt, Odej Kao

OpenIncrement introduces a unified framework addressing deep class-incremental learning and open set recognition by preserving feature relations during successive training. The approach leverages supervised contrastive learning and relation-based knowledge distillation to effectively detect novel samples and classify known ones, mitigating feature distortion. The method achieved an AUROC of 72.23% for open set recognition and 63.73% for inlier classification on CIFAR-100.

Key insights

  • Incremental learning methods significantly increase feature distortion, evidenced by higher R_s metric values (ratio of intra-class spread to inter-class spread), making features unsuitable for distance-based OSR.
  • Supervised contrastive learning effectively mitigates feature distortion, consistently yielding lower R_s values and maintaining well-separated, compact feature representations crucial for OSR.
  • Combining supervised contrastive learning with relation-based knowledge distillation allows a single model to both adapt to new classes and discern novel, unknown data while minimizing catastrophic forgetting.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.