Causal Intervention-based Counterfactual Generative Network for Generalized Zero-shot Learning

Generalized Zero-shot learning (GZSL) aims to recognize samples of both seen and unseen classes, while only annotated samples of seen classes are provided in training stage. Recently, generative methods have shown promising potency on GZSL task, which attempt to synthesis samples of unseen categories and thus translate the GZSL task into a supervised one. However, existing generative GZSL methods always generate samples from random noises and perform classification on original sample space, whose authenticity and discriminability is still unsatisfactory. In this paper, we propose a Causal Intervention-based Counterfactual Generative Network (CICGNet) for GZSL, which contains a counterfactual generative module (CGM) and a causal intervention classification module (CICM). The CGM is proposed to generate authentic synthesized samples based on counterfactual strategy. It generates diverse samples of each category via integrating specific semantic information and general context information. The CICM is designed to learn a discriminative classification space via exploring essential causalities among samples and categories by means of causal intervention. Comprehensive experiments conducted on five benchmark datasets demonstrate the competitive performance of our proposed method.