A model of hippocampal replay driven by experience and environmental structure facilitates spatial learning

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Neuronal replay in the hippocampus during rest and sleep is crucial for learning and memory.
• Replay sequences typically adhere to spatial constraints but can also generate novel sequences not based on prior behavior.

Purpose of the Study:

• To propose a novel stochastic replay mechanism for neural sequences.
• To investigate how prioritizing experiences based on strength, similarity, and inhibition of return impacts learning.

Main Methods:

• Developed a prioritized replay mechanism incorporating experience strength, similarity, and inhibition of return.
• Trained reinforcement learning agents using this mechanism and compared performance against random replay and a state-of-the-art algorithm.

Main Results:

• The prioritized replay mechanism significantly improved reinforcement learning agent performance compared to random replay.
• The model's performance approached that of a computationally intensive state-of-the-art algorithm.
• The stochastic nature of the mechanism and experience-dependent variables generated diverse replay statistics.

Conclusions:

• A unified, stochastic replay mechanism can generate diverse replay patterns.
• This mechanism is efficient in driving spatial learning and improving memory consolidation.
• The model offers a computationally efficient alternative for understanding and replicating neural replay phenomena.