A Contrastive Method for Continual Generalized Zero-Shot Learning

Chen Liang, Wentao Fan, Xin Liu, Shu-Juan Peng

Abstract: Generalized zero-shot learning (GZSL) aims to train a model that can classify seen and unseen samples based on semantic information. Continual learning, as one of the factors distinguishing artificial intelligence, has received more and more attention in recent years. In this paper, we propose a deep learning model based on a conditional generative model and a contrastive learning framework that can continuously learn from incoming data. In the training phase, knowledge distillation and generative replay mechanisms are used to accumulate past knowledge. In the testing phase, the test samples are mapped to the embedding space to perform generalized zero-shot classification task. Our model not only does not require additional buffers to store data so that it will not cause data leakage problems, but also through experiments show that our model can achieve state-of-the-art results on three benchmark datasets. The source code of the proposed model is available at: https://github.com/liangchen976/CCGZSL.