Semantic Intervention GAN for Generalized Zero-Shot Learning

In generalized zero-shot learning (GZSL), the method of transforming the zero-shot learning problem into a traditional classification problem is mostly handled by generating invisible data with the generative adversarial network (GAN). Currently, the data generated by the GAN has less differentiation between categories and the similarity within categories. In this paper, we modify CoGAN and optimize the loss function accordingly, and propose a semantically differentiated GAN structure, so that GAN has more inter-class differences and intra-class similarities when generating data, we call it S-CoGAN. We perform validation on MNIST and KDD99 datasets, demonstrating the effectiveness of our proposed network structure.