Similar concept matters: Prototype analogy for few-shot classification

Abstract

Few-shot classification aims to produce a classifier to recognize novel classes not seen during training with few labeled samples. The unseen classes and scarce samples make few-shot classification truly challenging. Transfer learning has been demonstrated to be an efficient paradigm for few-shot classification problem in recent literature. However, most methods based on transfer learning only utilize the parameters of the pre-trained model, ignoring the base features themselves which can be seen as the semantic concepts learned by the pre-trained model. In this paper, our main innovation lies in highlighting the importance of the learned semantic concepts of base classes. We propose a simple yet effective approach for few-shot classification to explore the pre-trained semantic features of base classes. Our approach innovatively employs prototype analogy inside and outside the few-shot classification task, to perform clustering and to select base features respectively, in an alternate and iterative way. We further design the best arrangement for these two steps. The initial centroids for clustering are constantly optimized by more and more accurate base features which are selected by the clustered novel prototypes from the previous iteration. When the iteration converges, the best semantic base features are selected to complete the prototypes of novel classes. Extensive experiments on four standard datasets and two deep backbones are conducted to demonstrate the effectiveness of our proposed prototype analogy method. Notably, our method requires neither sophisticated transductive algorithm nor additional learnable parameters besides the pre-trained model yet achieving comparable or even state-of-the-art performance on the miniImageNet, tieredImageNet and CIFAR-FS datasets.