Meta-hallucinating prototype for few-shot learning promotion

An effective way for few-shot learning (FSL) is to establish a metric space where the distance between a query and the prototype of each class is computed for classification, and the key lies on hallucinating the appropriate prototypes for each class of the given FSL task. Most existing prototypical approaches hallucinate the class-wise prototype based on the given support samples with an equal contribution assumption, i.e., each support sample contributes equally to the corresponding prototype. However, due to the limited-data regime as well as the strict assumption, the hallucinated prototypes often deviate from the ideal ones that are determined by the sample distribution of each unseen class, and thus causing poor generalization performance. To mitigate this problem, we present a prototype meta-hallucination approach which shows two aspects of advantages. On one hand, instead of directly inferring the complicated sample distribution, it meta-learns to establish a difference distribution based generative model that infers the distribution of inter-sample difference and synthesizes new labeled samples through fusing the sampled inter-sample difference and each given support sample. This empowers us to augment the support set with more content-diverse samples and is beneficial to reduce the bias in prototype hallucination. On the other hand, we argue that each support sample may contribute no-equally to the ideal prototype that it belongs to and their relations vary with class characteristics. Following this, our approach meta-learns to dynamically re-weight all support samples in prototype hallucination, which makes it flexible to locate the ideal prototype for each unseen class based on its characteristics. Experiments on four FSL benchmark datasets show that our approach can effectively improve the performance of the prototypical baseline and outperform several state-of-the-art competitors with a clear margin.