Reproduce “Adaptive Task Sampling for Meta-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
Chenghao Liu, Zhihao Wang, Doyen Sahoo, Yuan Fang, Kun Zhang, Steven C.H. Hoi
Researchers at Singapore Management University and Salesforce Research Asia developed Greedy Class-Pair Based Sampling (gcp-sampling), an adaptive task sampling mechanism that significantly improves the generalization performance of meta-learning algorithms for few-shot classification. This approach consistently enhanced classification accuracy across diverse meta-learning methods and datasets, yielding up to a 1.84% improvement in 1-shot accuracy on miniImageNet.
Key insights
- Adaptive task sampling, particularly focusing on 'hard' class pairs, is crucial for improving meta-learning generalization, outperforming uniform and class-based sampling.
- Task difficulty is context-dependent and best captured by the interaction between classes (i.e., class-pair relationships), rather than by individual class difficulty in isolation.
- The proposed gcp-sampling algorithm is model-agnostic, providing performance benefits across both optimization-based and metric-based meta-learning algorithms.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.