Reproduce “Cross-domain Few-shot Learning with Task-specific Adapters” 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
Wei-Hong Li, Xialei Liu, Hakan Bilen
Researchers at the University of Edinburgh's VICO Group developed a cross-domain few-shot learning approach that directly optimizes light-weight task-specific adapters on the support set during meta-test. This strategy achieved state-of-the-art performance on the Meta-Dataset benchmark, with an average accuracy increase of 7.5% on unseen datasets compared to previous methods.
Key insights
- Directly optimizing task-specific adapter weights on the support set is more robust and effective for cross-domain generalization than using an auxiliary network to predict them.
- Residual connections for adapters consistently outperform serial connections, allowing more stable and robust information flow while reducing the learning burden for new tasks.
- Including a pre-classifier affine transformation in the adaptation process further boosts performance by effectively aligning adapted features for the final nearest centroid classifier.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.