Meta-Generating Deep Attentive Metric for Few-Shot Classification

Learning to generate a task-aware base learner proves promising direction deal with few-shot learning (FSL) problem. Existing methods mainly focus on generating an embedding model utilized fixed metric (e.g., cosine distance) for nearest neighbour classification or directly linear classifier. However, due the limited discriminative capacity of such simple classifier, these fail generalize challenging cases appropriately. To mitigate this problem, we present novel deep meta-generation method that turns orthogonal direction, i.e., adaptively specific new FSL task based description few labelled samples). In study, structure using three-layers attentive network is flexible enough produce each task. Moreover, different from existing utilize uni-modal weight distribution conditioned samples generation, proposed meta-learner establishes multi-modal cross-class sample pairs tailored variational autoencoder, which can separately capture inter-class discrepancy statistics class and jointly embed all classes into generation. By doing this, generated be appropriately adapted pleasing generalization performance. demonstrate test three benchmark datasets gain competitive results state-of-the-art competitors.