Hippocampo-neocortical interaction as compressive retrieval-augmented generation.

Area of Science:

• Cognitive Neuroscience
• Computational Modeling
• Memory Systems

Background:

• The interplay between episodic (hippocampal) and semantic (neocortical) memory systems is crucial for learning, memory, and problem-solving.
• Neural mechanisms underlying this interaction remain largely unclear, hindering a comprehensive understanding of memory consolidation and retrieval.

Purpose of the Study:

• To present a computational model elucidating the neural mechanisms of hippocampal-neocortical interaction in memory and problem-solving.
• To simulate memory processes including encoding, recall, and generalization using a retrieval-augmented generation framework.

Main Methods:

• Developed a computational model simulating hippocampal encoding of compressed sequential experiences.
• Trained a neocortical generative network using replayed hippocampal data to capture gist and extract generalizable statistical patterns.
• Simulated retrieval-augmented generation, incorporating memory compression and consolidation mechanisms.

Main Results:

• The model demonstrates how the hippocampus retrieves episodic details to augment neocortical general knowledge for generation.
• It explains memory changes over time, including schema-based distortions, and shows how episodic and semantic memory contribute to problem-solving.
• The system enables efficient reconstruction of past events and prediction of future scenarios.

Conclusions:

• The proposed computational model provides a mechanistic account of episodic and semantic memory interactions.
• Retrieval-augmented generation offers a framework for understanding memory consolidation, generalization, and problem-solving capabilities.
• This interaction is vital for flexible cognitive functions, bridging specific experiences with generalized knowledge.