Reproduce “XCon: Learning with Experts for Fine-grained Category Discovery” 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
Yixin Fei, Zhongkai Zhao, Siwei Yang, Bingchen Zhao
The approach leverages dataset partitioning based on coarse self-supervised features and dual-level contrastive learning to enhance fine-grained generalized category discovery. XCon achieves state-of-the-art performance on four challenging fine-grained datasets and generic benchmarks, demonstrating accuracy improvements of up to 1.5% over previous methods.
Key insights
- Coarse self-supervised features can be effectively used to create data partitions that isolate and minimize class-irrelevant variations, enabling more focused fine-grained learning.
- Introducing separate 'expert projectors' for each sub-dataset allows the model to learn distinct, highly discriminative features for subtle differences within groups that are already coarsely similar.
- Combining a global coarse-grained learning objective with local fine-grained objectives leads to more robust and discriminative representations for GCD tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.