The hippocampal formation as a hierarchical generative model supporting generative replay and continual learning

Area of Science:

• Computational neuroscience
• Cognitive neuroscience
• Neuroscience

Background:

• The hippocampal formation's role in spatial navigation and episodic memory is complex.
• Existing frameworks struggle to explain flexible and prospective aspects of hippocampal dynamics.

Purpose of the Study:

• To propose a novel computational theory of the hippocampal formation as a hierarchical generative model.
• To explain how sequential experiences are organized into spatiotemporal contexts.

Main Methods:

• Developed a hierarchical generative model of the hippocampal formation.
• Simulated the model's ability to organize sequential experiences, such as spatial navigation trajectories.
• Characterized hippocampal reactivations as generative replay for continual learning.

Main Results:

• The model successfully learns and retains multiple spatial navigation trajectories by organizing them into spatial maps.
• The model reproduces flexible and prospective aspects of hippocampal dynamics.
• The theory reconciles multiple proposed roles of the hippocampal formation.

Conclusions:

• The hippocampal formation functions as a hierarchical generative model organizing experiences into spatiotemporal contexts.
• Generative replay supports continual learning of multiple sequential experiences.
• This theory integrates map-based navigation, episodic memory, and imagination within a unified framework.