Inductive Unsupervised Domain Adaptation for Few-Shot Classification via Clustering

Few-shot classification tends to struggle when it needs adapt diverse domains. Due the non-overlapping label space between domains, performance of conventional domain adaptation is limited. Previous work tackles problem in a transductive manner, by assuming access full set test data, which too restrictive for many real-world applications. In this paper, we out tackle issue introducing inductive framework, DaFeC, improve Domain via Clustering. We first build representation extractor derive features unlabeled data from target (no necessary) and then group them with cluster miner. The generated pseudo-labeled labeled source-domain are used as supervision update parameters few-shot classifier. order high-quality pseudo labels, propose Clustering Promotion Mechanism, learn better Similarity Entropy Minimization Adversarial Distribution Alignment, combined Cosine Annealing Strategy. Experiments performed on FewRel 2.0 dataset. Our approach outperforms previous absolute gains (in accuracy) 4.95%, 9.55%, 3.99% 11.62%, respectively, under four settings.