A generative model of memory construction and consolidation.

Area of Science:

• Cognitive Neuroscience
• Computational Psychiatry
• Neuroscience

Background:

• Episodic memories are reconstructed, not passively stored, and share neural bases with imagination.
• Memory consolidation leads to schema-based distortions, like boundary extension, increasing with time.
• Existing models struggle to fully integrate sensory details with conceptual schemas.

Purpose of the Study:

• To present a novel computational model of memory (re)construction and consolidation.
• To explain how episodic memories combine unique sensory features with schema-based predictions.
• To account for memory distortions and the interplay between hippocampal and neocortical systems.

Main Methods:

• Developed a computational model where hippocampal replay trains generative models (variational autoencoders).
• Simulated memory reconstruction from latent variable representations in key brain regions (entorhinal, medial prefrontal, temporal cortices, hippocampal formation).
• Tested model predictions against empirical data on memory age and hippocampal lesions.

Main Results:

• The model replicates effects of memory age and hippocampal lesions observed in previous studies.
• It provides mechanisms for semantic memory, imagination, episodic future thinking, and relational inference.
• It explains schema-based distortions, including boundary extension, by combining unique and predictable memory elements.

Conclusions:

• Hippocampal replay training generative models offers a comprehensive framework for memory construction and imagination.
• The model elucidates the efficient integration of hippocampal and neocortical functions in memory.
• It highlights how memory systems optimize storage for novel information while incorporating conceptual schemas.