Data-Free Continual Learning in a Spiking Network via Targeted Sleep Consolidation and Protection-Aware Normalization: A Pre-registered Evaluation

Neural networks forget earlier tasks when trained on new ones. Regularization methods such as Elastic Weight Consolidation (EWC) reduce forgetting at the cost of plasticity, and replay methods store past data. We study a brain-inspired spiking network that learns a sequence of lessons with spike-timing-dependent plasticity and stores no training data. After each lesson an offline “sleep” phase consolidates only the synapses leaving each stored memory’s input cells, and a protection-aware synaptic normalization excludes the protected part of every synapse from competition. Every confirmatory hypothesis was pre-registered in version control and tested on held-out random seeds. On five sequential digit pairs of MNIST, the method reached 54.7% ten-class accuracy versus 36.4% for EWC whose strength was tuned on MNIST itself (+18.3 points, 10/10 seeds); the advantage held under ten random task orders (+17.1, 10/10) and for a variant that selects memories from lesson activity without labels. Removing protection-aware normalization alone reduced accuracy by 22.1 points. On Fashion-MNIST, with no re-tuning of our method, the advantage over a re-tuned EWC persisted but was much smaller (+4.5 points, 9/10 seeds), and the label-free variant was the best method (51.6%). Absolute accuracy remains below a raw-pixel nearest-class-mean reference (73.5% and 62.2%), so our claims concern the relative behaviour of sequential learners under identical constraints. We report negative results, including an apparent theta-rhythm benefit that a matched control traced to extra learning time.