Zero-Shot Incremental Learning Algorithm Based on Bi-alignment Mechanism

Zero-shot incremental learning aims to enable the model to generalize to new classes without forgetting previously learned classes. However, the problem of the semantic gap between old and new sample classes has been puzzling to researchers. Therefore, this paper proposes a zero-shot incremental learning algorithm based on a bi-alignment mechanism, called BANet, which is mainly divided into an intra-class alignment module Intra-CA and an inter-class alignment module Inter-CA. The model can better extract image features, improve image classification accuracy, and fundamentally alleviate the catastrophic forgetting of the network. Extensive experiments have been conducted on two basic datasets, CUB-200-2011 and CIFAR100, and the experimental results show that our proposed algorithm outperforms the current state-of-the-art incremental learning algorithms.