Episode Adaptive Embedding Networks for Few-Shot Learning

Few-shot learning aims to learn a classifier using few labelled instances for each class. Metric-learning approaches few-shot embed into high-dimensional space and conduct classification based on distances among instance embeddings. However, such embeddings are usually shared across all episodes thus lack the discriminative power generalize classifiers according episode-specific features. In this paper, we propose novel approach, namely \emph{Episode Adaptive Embedding Network} (EAEN), of instances. By leveraging probability distributions in an episode at channel-pixel embedding dimension, EAEN can not only alleviate overfitting issue encountered tasks, but also capture features specific episode. To empirically verify effectiveness robustness EAEN, have conducted extensive experiments three widely used benchmark datasets, under various combinations different generic backbones classifiers. The results show that significantly improves accuracy about $10\%$ $20\%$ settings over state-of-the-art methods.