Class-Incremental Generalized Zero-Shot Learning

Zero-Shot Learning (ZSL) focuses on transferring knowledge learned from the source domain to the target domain. In the classic setting, test data only come from the target domain. Recently, a more reasonable setting GeneralizedZSL (G-ZSL) evaluates the model by recognizing instances from both the source and the target domain. However, G-ZSL still have some unrealistic assumptions. One of such hypotheses is the class-fixed setting: during the training time, G-ZSL algorithms only learn to recognize a fixed class set, and we cannot modify the target model unless retraining. The capacity of learning continuously and efficiently is crucial for learning algorithms in a real scenario. In this paper, we extend G-ZSL to a more realistic setting: instead of supposing the class-fixed training strategy, the incoming data come in the way of class-incremental order. In different learning episodes, disjoint groups of categories are utilized to train the G-ZSL models. As the training process goes on, the source domain is expanding. In such a Class-Incremental Generalized Zero-Shot Learning (CIG-ZSL) setting, learning algorithms are expected to not only transfer the learned knowledge to the target domain but also to remember the previously learned knowledge. We propose a Dual Path Learner (DPL) algorithm to validate the possibility of solving the CIG-ZSL task. Experiments on several benchmarks show that DPL has the capacity of remembering the knowledge learned from previous source instances and be able to migrate all the knowledge to the target domain as the expanding of the source domain.