The Role of Neural Replay in Shaping and Utilizing Neural Representations Supporting Learning, Memory, and Decision-Making

Abstract

Flexible generalization depends on internal representations that capture the latent structure of the environment and can be dynamically updated as that structure changes. This thesis investigates the role of neural replay, the offline reactivation of sequential neural patterns, in building, refining, and leveraging such representations across wakefulness and sleep.

The first empirical project combined functional magnetic resonance imaging (fMRI) with computational modelling to study how replay supports structure learning and generalization. Fifty-two participants performed a sequential decision task in which multiple options shared hidden reward sources. A latent cause inference model captured participants' discovery of this structure and outperformed standard reinforcement learning models. Backward replay in visual cortex selectively targeted unobserved sequences sharing the inferred latent reward structure, increased as participants discovered that structure, and tracked trial-by-trial non-local value updating from the computational model. When the latent structure changed on a second day, replay reorganised accordingly, including a directional shift for newly generalizable paths.

The second empirical project examined how sleep transforms the neural basis of sequential memory retrieval. Participants learned hierarchically organized sequences and were tested before and after overnight sleep. Sequential reactivation during retrieval shifted in regional profile across consolidation: visual cortex dominated before sleep, whereas parietal cortex emerged as the primary locus after sleep, with replay strength predicting retrieval accuracy. This reorganisation is consistent with the systems consolidation hypothesis that sleep drives the redistribution of memory representations toward cortical networks.

A methodological book chapter situated these findings within the broader landscape of tools for measuring neural plasticity in humans. Together, the three projects converge on a view of replay as an active computational mechanism sampling from an internal generative model supporting multiple functions including rapid value propagation over inferred structure during wakefulness, and driving cortical integration of memory representations during sleep.