Rehearsal revealed: The limits and merits of revisiting samples in continual learning

Learning from non-stationary data streams and overcoming catastrophic\nforgetting still poses a serious challenge for machine learning research.\nRather than aiming to improve state-of-the-art, in this work we provide insight\ninto the limits and merits of rehearsal, one of continual learning's most\nestablished methods. We hypothesize that models trained sequentially with\nrehearsal tend to stay in the same low-loss region after a task has finished,\nbut are at risk of overfitting on its sample memory, hence harming\ngeneralization. We provide both conceptual and strong empirical evidence on\nthree benchmarks for both behaviors, bringing novel insights into the dynamics\nof rehearsal and continual learning in general. Finally, we interpret important\ncontinual learning works in the light of our findings, allowing for a deeper\nunderstanding of their successes.\n