OntheEffectivenessofLipschitz-Driven RehearsalinContinualLearning

Rehearsal approaches enjoy immense popularity with Continual Learning (CL) practitioners. These methods collect samples from previously encountered data distributions in a small memory buffer; subsequently, they repeatedly optimize on the latter to prevent catastrophic forgetting. This work draws attention to a hidden pitfallofthis widespread practice: repeated optimization onasmall pool of data inevitably leads to tight and unstable decision boundaries, which are a major hindrance to generalization.