<ul><li>Flashback Learning (FL) is a new method designed to balance stability and plasticity in Continual Learning (CL).</li><li>FL differs from previous approaches by bidirectionally regularizing model updates to incorporate new knowledge while retaining old knowledge.</li><li>It operates through a two-phase training process and can be integrated into various CL methods, showing improvements over baseline methods.</li><li>Empirical results demonstrate up to 4.91% accuracy improvement in Class-Incremental and 3.51% in Task-Incremental settings, surpassing state-of-the-art methods on datasets like ImageNet.</li></ul>

Flashbacks to Harmonize Stability and Plasticity in Continual Learning

Discover more