One-Stage Training Generative Paradigm for Generalized Zero-Shot Learning

Abstract

Zero-shot learning image classification aims to identify unseen classes not present during training. Generalized zero-shot learning (GZSL) is more in line with realistic scenarios due to its ability of recognizing both seen and unseen classes. Current GZSL methods mostly utilize generative adversarial networks (GANs) but typically follow a two-stage training: first, training the GAN and then, using its synthetic features to train a classifier, which is limited by isolated optimizations rather than federated. We propose a novel One-stage Training Generative Paradigm that incorporates the classifier as a unique synthetic label generator and builds a three-player game involving a generator, discriminator, and classifier, which ensures a unified optimization objective, eliminating the discrete optimization approach of two-stage methods. We also propose a label-attribute classifier that leverages both labels and attributes, surpassing traditional softmax classifiers that only use labels. Our test results show the effectiveness of the proposed methods.