HMRM: Hierarchy-aware Misclassification Risk Minimization for few-shot learning

Hierarchical Few-shot Learning (HFSL) is a practical research of recognizing new categories from insufficient samples, which leverages multi-grained knowledge among samples to solve the data limitation problem. Current HFSL methods have delivered exceptional improvement and obtained outstanding performance. However, there are two potential and inevitable misclassification risk problems: Indifferent Risk occurs since most HFSL only concerns the correct classification of query samples and ignores the misclassifications; Indiscriminate Risk emerges in case of not distinguishing the error degree of these misclassifications. In this paper, we utilize multi-grained knowledge to propose Hierarchy-aware Misclassification Risk Minimization (HMRM) algorithm to minimize aforementioned two risks. HMRM consists of Cross Fine-grained Indifferent Risk Minimization (CFIRM) and Cross Coarse-grained Indiscriminate Risk Minimization (CCIRM) submodules. First, we propose CFIRM to design a fine-grained contrastive loss, which considers and evaluates misclassification results with fine-grained knowledge. Second, we present CCIRM to assign coarse-grained cost for misclassification via coarse-grained knowledge and cost-sensitive learning. Experimental results on FC100, CUB-200-2011, CIFAR-FS, and miniImageNet indicate the effectiveness of HMRM method. For instance, the accuracy of HMRM is at least 1.10% and 1.30% better than other methods in 5-way 1-shot on FC100 and CIFAR-FS. Our code is available at https://github.com/fhqxa/HMRM.