Circle Loss Based Metric Learning For Few-Shot Classification

Abstract

Few-shot classification aims to identify new categories from only a handful of labeled examples and has drawn considerable attention in machine learning. An efficient and effective approach for solving few-shot classification has featured episodic training based metric learning which attempts to learn a task-generic embedding space on extremely large episodes, each with a small labeled support set and its corresponding query set. Distances between embeddings of examples from the query set and support set are then used as a notion of similarity to do classification. This paper first adds a margin between different classes to increase the inter-class distance in the embedding space(we call this method ML-FSC). Furthermore, considering that a similarity score that deviates far from the optimum should be emphasized, this paper introduces the circle loss into the episodic training based metric learning, which simply re-weights each similarity to highlight the less-optimized similarity scores, contributing to a more discriminative embedding learning(we call this method CL-FSC). We conduct comprehensive experiments to validate our method and we set new state-of-the-art performance on three popular few-shot classification benchmarks, namely omniglot, miniImageNet and tieredImageNet.