Class-Incremental Continual Learning into the eXtended DER-verse

• 10.1109/TPAMI.2022.3206549

• September 2022

• IEEE TPAMI

• Continual Learning Distillation Rehearsal Class-Incremental

The study focuses on Class-Incremental Continual Learning, aiming to address the challenge of catastrophic forgetting in Deep Networks. The researchers improve upon Dark Experience Replay (DER), with a new method called eXtended-DER (X-DER). X-DER revises its replay memory to incorporate new information and facilitate learning of previously unseen classes. Results show that X-DER outperforms existing methods on standard benchmarks like CIFAR-100 and miniImageNet, as well as a new benchmark introduced in the study. Ablation studies highlight the importance of Knowledge Distillation and the role of flatter minima in continual learning setups.