SPARCL: Spectral Partitioned Analytic Continual Learning
The paper pins analytic continual-learning drift on “spectral interference,” not gradient overwriting.
SPARCL keeps a high-energy core of the running autocorrelation fixed for old classes, then updates only the residual space with closed-form recursive least squares. The authors claim this preserves the core contribution to old logits while still learning incoming classes. They report results on CIFAR-100, CUB-200, ImageNet-R, and ImageNet-A with frozen ViT-B/16 features, narrowing much of the gap to stronger representation-matching methods. ArXiv · AI/CL/LG's note
SPARCL keeps a high-energy core of the running autocorrelation fixed for old classes, then updates only the residual space with closed-form recursive least squares. The authors claim this preserves the core contribution to old logits while still learning incoming classes. They report results on CIFAR-100, CUB-200, ImageNet-R, and ImageNet-A with frozen ViT-B/16 features, narrowing much of the gap to stronger representation-matching methods. ArXiv · AI/CL/LG's note
score 4