CSN: Component supervised network for few-shot classification

Abstract

The few-shot classification (FSC) task aims to classify data with limited labeled examples across different categories. Typically, researchers pre-train a feature extractor using base data and use it to extract features from novel data for classification Notably, the novel set only has a few annotated samples and has non-overlapped categories from the base set, which leads to the fact that the pre-trained feature extractor cannot adapt to the novel data flawlessly. We dub this problem as Feature-Extractor-Maladaptive (FEM) problem. Starting from the root cause of this problem, this paper presents a new scheme called Component Supervised Network (CSN), to improve the performance of FSC. We believe that even though the categories in the base and novel sets are different, their sample components share similarities. For example, both cats and dogs contain leg and head components. These entity components are intra-class stable and have cross-category versatility, making them useful for generalization to new categories. However, finding common information among different categories is difficult in real-world scenarios, hindering the possibility of modeling based on this assumption. To overcome this, we first design a Dictionary-based Implicit-Component Generator (DICG) to mine common information of different sets; then construct an implicit-component-based auxiliary task to improve the adaptability of the feature extractor. We evaluated our CSN on three benchmark datasets (mini-ImageNet, tiered-ImageNet, and FC100) and achieved improvements of at least 0.9% compared to classical methods, demonstrating the efficiency of our approach.