Abstract
During sleep and rest, the brain engages in internally generated sequences of activity which reflect encoding of recent experiences. This replay of neural activity is believed to promote memory consolidation, preferentially reinforcing the memories encoded by the replayed activity. It is thought that preferential replay of some activity can optimise the processing and retention of new information to optimise future behaviour, but it remains unclear which experiences are prioritised for replay, or which features of an awake experience influence replay prioritisation. Replay depends on the hippocampus, a brain structure heavily involved in forming new memories, and recruits a wide range of other brain areas to enable systems-level consolidation coordinated across the brain.This thesis presents a combination of computational modelling and in vivo behavioural and electrophysiological experiments used to investigate how reward and non-reward experiences influence replay. I developed and implemented a maze-based reinforcement learning task with stochastic rewards, in which both rewarded and unrewarded trials are informative for learning.
Computational modelling based on a reinforcement learning framework showed that biasing replay by reward or reward-prediction error can enhance learning, but replay of a range of trials (rewarded and unrewarded) is necessary. Modelling of rats’ behaviour on the same task suggested that they preferentially replayed experiences which generated high reward-prediction errors between training sessions. Preliminary multi-unit recordings made from the hippocampus, which is heavily implicated in replay, and the nucleus accumbens, which responds to reward and receives input from the hippocampus, suggests that neural activity encoding spatial and reward information in these two structures is replayed during post-task rest. This is a likely mechanism by which reinforcement learning can occur after behaviour has taken place.
The work presented in this thesis extends the understanding of how reward influences learning, memory consolidation and replay, particularly in offering evidence for the biasing effect of reward-prediction error.
| Date of Award | 24 Mar 2020 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Matt W Jones (Supervisor) & Nathan F Lepora (Supervisor) |