Exploring the roles of memory replay in targeted memory reactivation and birdsong development: Insights from computational models of complementary learning systems

Abstract Replay facilitates memory consolidation in both biological and artificial systems. Using the complementary learning systems (CLS) framework, we study replay in both humans and birds through computational modelling. We investigate impacts of replay triggered by targeted memory reactivation during sleep and experiments examining how sleep affects the development of birdsong in young songbirds. We show that qualitatively realistic sleep effects can be captured by highly abstracted, idealised CLS models. Our modelling sheds theoretical insights on the mechanisms underlying both strengthening and weakening effects of targeted memory reactivation, and supports the empirical hypothesis that replay drives overnight performance deterioration and correlates positively with the final performance in birdsong development. Author summary Taking a computational approach, we investigated the roles of memory replay in two complementary learning systems (CLS) models capturing realistic sleep effects observed in two real-life experiments on targeted memory reactivation (TMR) and birdsong development respectively. Our two CLS models are abstract and identical in architecture, and they are distinct in terms of where replay is generated. While the TMR model produces replay samples using its hippocampus, the birdsong model does so using the sensorimotor cortex. We found that certain TMR effects could characterise different TMR models, which might account for individual differences in human subjects. The results of the birdsong model support the idea that the dramatic overnight oscillations in performance accuracy which are observed during birdsong development are mainly driven by memory replay, and that long-term performance gain can be achieved despite short-term performance deterioration during the early nights of development. As we studied the two experiments using the unified CLS framework, we discuss how replay contributes to sleep-dependent performance changes from the perspective of systems consolidation.