Incremental Embedding Learning via Zero-Shot Translation

Modern deep learning methods have achieved great success in machine and computer vision fields by a set of pre-defined datasets. Howerver, these perform unsatisfactorily when applied into real-world situations. The reason this phenomenon is that new tasks leads the trained model quickly forget knowledge old tasks, which referred to as catastrophic forgetting. Current state-of-the-art incremental tackle forgetting problem traditional classification networks ignore existing embedding networks, are basic for image retrieval, face recognition, zero-shot learning, etc. Different from semantic gap between spaces two adjacent main challenge under setting. Thus, we propose novel class-incremental method network, named translation (ZSTCI), leverages estimate compensate without any exemplars. Then, try learn unified representation sequential process, captures relationships previous classes current precisely. In addition, ZSTCI can easily be combined with regularization-based further improve performance networks. We conduct extensive experiments on CUB-200-2011 CIFAR100, experiment results prove effectiveness our method. code has been released https://github.com/Drkun/ZSTCI.