Replay as a Basis for Backpropagation Through Time in the Brain

Area of Science:

• Neuroscience
• Computational Neuroscience
• Artificial Intelligence

Background:

• Episodic memory formation is a key neuroscience challenge.
• The hippocampus, crucial for episodic learning, exhibits recurrent connectivity and offline replay.
• The precise function of hippocampal replay events remains debated.

Purpose of the Study:

• To propose a biologically plausible model for episodic learning leveraging offline replay.
• To introduce a novel variant of backpropagation through time (BPTT) suitable for neural networks.
• To explain the functional significance of hippocampal replay events in memory consolidation.

Main Methods:

• Developed a reversible recurrent neural network (R2N2) model.
• Implemented a biologically plausible variant of BPTT within R2N2.
• Utilized forward and backward offline replay for information transfer between cache and consolidator networks.
• Tested R2N2 on computer science benchmarks and simulated the rodent delayed alternation T-maze task.

Main Results:

• R2N2 successfully models episodic learning using offline replay.
• The model demonstrates one-shot learning (cache) and statistical learning (consolidator).
• R2N2 outperforms existing methods like random feedback local online learning and reservoir networks.
• The model's architecture eliminates the need for artificial external memory stores, unlike standard BPTT.

Conclusions:

• R2N2 provides a biologically plausible mechanism for episodic memory formation.
• The model elucidates the functional role of hippocampal replay in memory processes.
• R2N2 offers a promising computational framework for understanding memory and learning in the brain.