Artificial neural networks (ANNs) suffer from catastrophic forgetting—when learning multiple tasks sequentially, they perform well on the most recent task at the expense of previously learned tasks. This study tests the hypothesis that implementing a sleep-like phase in ANNs can alleviate this. Sleep was implemented as offline training with local unsupervised Hebbian plasticity rules and noisy input. Results show that this 'Sleep Replay Consolidation' (SRC) algorithm recovered old tasks otherwise forgotten, replaying memories spontaneously during the sleep phase. Increased representational sparseness and neuronal activity related to old tasks, while decreasing new task activity, suggest that sleep-like dynamics can mitigate catastrophic forgetting.

Timothy Tadros, Giri P. Krishnan, Ramyaa Ramyaa, Maxim Bazhenov