Many-Class Few-Shot Learning on Multi-Granularity Class Hierarchy

We study many-class few-shot (MCFS) problem in both supervised learning and meta-learning settings. Compared to the well-studied many-shot few-class problems, MCFS commonly occurs practical applications but has been rarely studied previous literature. It brings new challenges of distinguishing between many classes given only a few training samples per class. In this paper, we leverage class hierarchy as prior knowledge train coarse-to-fine classifier that can produce accurate predictions for The propose model, "memory-augmented hierarchical-classification network (MahiNet)", performs classification where each coarse cover multiple fine classes. Since it is challenging directly distinguish variety data class, MahiNet starts from over coarse-classes with more whose labels are much cheaper obtain. reduces searching range thus alleviates "many classes". On architecture, firstly deploys convolutional neural (CNN) extract features. then integrates memory-augmented attention module multi-layer perceptron (MLP) together probabilities While MLP extends linear classifier, KNN targeting "few-shot" problem. design several strategies meta-learning. addition, two novel benchmark datasets "mcfsImageNet" "mcfsOmniglot" specially designed experiments, show outperforms state-of-the-art models on problems