Semantic Grouping with Dual-Strategy Distributional Rehearsal for Continual Learning

Continual learning systems face catastrophic forgetting, where acquiring new knowledge degrades performance on prior tasks. We propose SGC, a novel exemplar-free framework addressing this through three key innovations: (1) a Max-Cut graph-based partitioning strategy that groups incoming classes into semantically coherent clusters; (2) a dual-strategy distributional rehearsal (DSDR) combining archetype interpolation and GMM sampling to generate high-fidelity synthetic features preserving past class distributions; and (3) a group-focused training regimen that updates only lightweight group-specific modules while freezing the shared backbone. Extensive experiments on CIFAR-100 and four fine-grained datasets (CUB-200, Flower, Stanford-Cars, Food-101) demonstrate superior performance over state- of-the-art exemplar-free methods.