NORACL: Neurogenesis for Oracle-free Resource-Adaptive Continual Learning

arXiv cs.AI · 6 hours ago

Researchers propose NORACL, a method that addresses the stability-plasticity trade-off in continual learning by growing neural networks adaptively rather than using fixed architectures. The approach monitors representational and plasticity saturation to add neurons only when needed, achieving accuracies comparable to or better than oracle-sized baselines while using fewer parameters. This adaptive growth enables networks to maintain previously learned knowledge while retaining capacity for new tasks, with different layer types expanding based on task relationships.