Task-aware prototype refinement for improved few-shot learning

In realistic scenarios, few-shot classification aims to generalize from common classes to novel classes with limited labeled samples. Most of existing transductive methods concentrate on probing into instance-prototype relations in a fixed way, without considering task-relevant information. In this paper, we perform task-aware prototype refinement (TAPR) explicitly for metric-based few-shot learning. Instead of utilizing fixed prior of queries, we adaptively estimate the query distribution, which can accommodate to both balanced and imbalanced situations. Concretely, on the basis of discriminative features from a holistic pre-training and pre-processing stage, we make novel attempts to make the best of task-aware and instance-aware knowledge to conduct selecting and denoising of samples for prototype generation and iterative rectification, which are complementary to each other. Extensive experimental results on four popular benchmark datasets (CUB, CIFAR-FS, miniImageNet and tieredImageNet) demonstrate that our TAPR outperforms most methods in inductive and balanced transductive settings. Besides, it achieves good generalization while maintaining high accuracy in the imbalanced and cross-domain setting.