Zero-Shot Learning with Superclasses

Abstract

Zero-shot learning (ZSL) can be regarded as transfer learning from seen classes to unseen ones so that the later can be recognized without any training samples. Its main difficulty lies in that there often exists a large domain gap between the seen and unseen class domains. Inspired by the fact that an unseen class is not strictly ‘zero-shot’ (thus easier to recognize) if it falls into a superclass that consists of one or more seen classes, we propose a new ZSL model, termed ZSL with superclasses (ZSLS), that leverages the superclasses as the bridge between seen and unseen classes to narrow the domain gap. By generating the superclasses with k-means clustering over all seen and unseen class prototypes, we formulate ZSLS as a min-min optimization problem. An efficient iterative algorithm is also developed for model optimization. Extensive experiments show that our model achieves the state-of-the-art results.