Transductive Prototypical Network For Few-Shot Classification

Few-shot learning is the key step towards human-level intelligence. Prototypical Network is a promising approach to address the key issue of over-fitting for few-shot learning. Nevertheless, the original Prototypical Network only uses one or few labeled instances to represent the corresponding class, which easily deviates from the real class distribution leading to the imprecise classification results. In this paper, we propose Transductive Prototypical Network (Td-PN), a universal transductive approach that refines the class representations by merging scarce labeled samples and high-confidence ones of target set. Our proposed Td-PN first maps the samples to a classifying-friendly (discriminative) embedding space by redesigning a weighted contrastive loss, then utilizes the transductive inference to obtain the powerful prototype representation for each class. Experiments demonstrate that our approach outperforms the state-of-the-art algorithms.