Fewmatch: Dynamic Prototype Refinement for Semi-Supervised Few-Shot Learning

Semi-Supervised Few-shot Learning (SS-FSL) investigates the benefit of incorporating unlabelled data in few-shot settings. Recent work has relied on the popular Semi-Supervised Learning (SSL) concept of iterative pseudo-labelling, yet often yield models that are susceptible to error propagation and are sensitive to initialisation. Alternative work utilises the concept of consistency regularisation (CR), a popular SSL state of the art technique where a student model is trained to consistently agree with teacher predictions under different input perturbations, without pseudo-label requirements. However, applications of CR to the SS-FSL set-up struggle to outperform pseudo-labelling approaches; limited available training data yields unreliable early stage predictions and requires fast convergence that is not amenable for, typically slower to converge, CR approaches. In this paper, we introduce a prototype-based approach for SS-FSL that exploits model consistency in a robust manner. Our Dynamic Prototype Refinement (DPR) approach is a novel training paradigm for few-shot model adaptation to new unseen classes, combining concepts from metric and meta-gradient based FSL methods. New class prototypes are alternatively refined 1) explicitly, using labelled and unlabelled data with high confidence class predictions and 2) implicitly, by model fine-tuning using a data selective CR loss. DPR affords CR convergence, with the explicit refinement providing an increasingly stronger initialisation. We demonstrate method efficacy and report extensive experiments on two competitive benchmarks; miniImageNet and tieredImageNet. The ability to effectively utilise and combine information from both labelled base-class and auxiliary unlabelled novel-class data results in significant accuracy improvements.