eLife Assessment
This paper provides an important proposal for why learning can be much faster and more accurate if synapses have a fast component that immediately corrects errors, as well as a slower component that corrects behavior averaged over a longer timescale. It is convincingly shown that integrating these two learning timescales improves performance compared to classical strategies, particularly in terms of robustness and generalization when learning new target signals. However, the biological plausibility and justification for the proposed rapid learning mechanism require further elaboration and supporting mechanistic examples.