An Efficient Class-incremental Learning Strategy with Frozen Weights and Pseudo Exemplars

In this paper, we propose a novel and efficient class-incremental learning approach that does not necessitate the storage of old data after training each task. The proposed approach uses the autoencoder's decoder to generate pseudo data to consolidate the model and sets a subset of relevant weights in the encoder layers to learn new knowledge while freezing most weights. It uses no extra storage space to save old data. The experimental results also show the performance of the proposed approach.