Bridging Few-Shot Learning and Adaptation: New Challenges of Support-Query Shift

Few-Shot Learning (FSL) algorithms have made substantial progress in learning novel concepts with just a handful of labelled data. To classify query instances from classes encountered at test-time, they only require support set composed few samples. FSL benchmarks commonly assume that those queries come the same distribution as set. However, realistic setting, data is plausibly subject to change, situation referred Distribution Shift (DS). The present work addresses new and challenging problem under Support/ Query (FSQS) i.e., when are sampled related but different distributions. Our contributions following. First, we release testbed for FSQS, including datasets, relevant baselines protocol rigorous reproducible evaluation. Second, observe well-established unsurprisingly suffer considerable drop accuracy facing stressing significance our study. Finally, show transductive can limit inopportune effect DS. In particular, study both role Batch-Normalization Optimal Transport (OT) aligning distributions, bridging Unsupervised Domain Adaptation FSL. This results method efficiently combines OT celebrated Prototypical Networks. We bring compelling experiments demonstrating advantage method. opens an exciting line research by providing strong baselines. code available https://github.com/ebennequin/meta-domain-shift.