Meta hyperbolic networks for zero-shot learning

Abstract

Zero-Shot Learning (ZSL) aims at generalizing the classification experience from seen classes to unseen classes with auxiliary side information, among which word vectors of class names and class attributes are popular ones. While word vectors technique is more practical for large datasets, its performance is usually slightly worse than that of attributes since it requires no human annotations. In this paper, we focus on the utilization of word vectors in ZSL by exploiting their hierarchical knowledge for large dataset and propose Meta Hyperbolic Networks (MHN). Specifically, we present Poincaré Graph Convolutional Networks (P-GCN), which transforms the word vectors into a Poincaré ball, and then encodes them with Poincaré Graph Convolutional layers. During training, the image classifiers from Convolutional Neural Networks (CNN) and P-GCN weights are aligned to ensure an accurate mapping between them. Moreover, we further develop a short-term memory episode learning on it to relieve the model’s inherent bias towards seen classes. Extensive experiments on the popular ImageNet dataset show the competitive performance for both ZSL and generalized ZSL.