Reproduce “Domain-Adaptive Few-Shot 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
An Zhao, Mingyu Ding, Zhiwu Lu, Tao Xiang, Yulei Niu, Jiechao Guan, Ji-Rong Wen, Ping Luo
Researchers introduce Domain-Adaptive Few-Shot Learning (DA-FSL) and propose the Domain-Adversarial Prototypical Network (DAPN) to learn novel categories from few examples in new domains. DAPN consistently outperforms baselines, achieving a 27.25% accuracy in a 5-way 1-shot miniImageNet task, by intelligently reconciling domain alignment and class discriminativeness.
Key insights
- Successfully reconciling the conflicting objectives of global domain alignment and maintaining per-class discriminativeness is crucial for DA-FSL, achieved by strategically applying domain confusion after feature embedding and domain discrimination before it.
- Introducing a domain discriminative loss (L_ds) on initial features encourages the network to learn domain-specific features, allowing subsequent global domain alignment (L_dc) without sacrificing critical class-specific information.
- Dynamically balancing multiple loss functions through uncertainty-based adaptive re-weighting is essential for stable and effective training in multi-objective optimization problems like DA-FSL.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.