More task-balanced Class-Incremental Learning

Although deep learning has achieved high performance on many tasks, neural networks suffer from catastrophic forgetting in incremental learning scenarios. Class-Incremental Learning focuses on the scenario where new category data continuously arrives. PASS (Prototype Augmentation and Self-Supervised) is a classic method that does not require storing old exemplars. Based on PASS, we propose three improvement measures from the perspective of making the new and old tasks more balanced. Firstly, we propose a balanced prototype replay (Balanced PR) strategy to make the image number of new and old categories that the model head encounters more balanced. Secondly, at the feature knowledge distillation (FKD) stage, we assign different weights to the incoming data based on the confidence of the old model, which is named Focal FKD. The goal is to minimize interference with learning new task data while preserving the representation capability of the model backbone on old data. Thirdly, we combine FKD with weight regularization on the model backbone. In the feature space, FKD is beneficial for preserving the positions of the old categories while weight regularization helps to separate the new and old categories well. Experiments on the benchmark indicate that there has been obvious improvement in the performance of PASS after our modifications.