Venue: Centre Broca

IMN
Team: Computational Neuroscience 

Thesis directed by Frédéric Alexandre

Title

Generative models in the neocortex and the hippocampus

Abstract

The neocortex and the hippocampus are complementary learning systems, supporting semantic and episodic memory respectively. The hippocampus quickly stores experiences, whereas the neocortex slowly learns regularities from experiences. After storage, the hippocampus replays memories to the neocortex for memory consolidation. More recently, hippocampal replay was found not only to recapitulate past experiences, but also to construct novel ones. Therefore, researchers have proposed that the hippocampus holds a generative model of the world and that replay is a generative process. The idea of a generative model in the brain has a long tradition, but it is originally associated with the neocortex, which inspired several famous generative models. How the hippocampus implements this construction system in coordination with the neocortex is still largely unknown. The objective of this thesis is to study how neural circuits in the neocortex and the hippocampus implement generative models of the world, enabling us to remember past experiences and imagine future ones. This question pushes us to examine the nature of generative models, their biological plausibility and compatibility with the neocortex and the hippocampus. We propose a neocortical-hippocampal network model, in which the neocortex forms a hierarchical predictive coding network which culminates in the entorhinal cortex. The entorhinal cortex is also the input of a hippocampal generative model, which is trained to reconstruct its inputs by learning compressed, sparse representations in CA1. The neocortical network learns the structure across experiences, allowing it to predict the missing information in its input, but without retaining the individual details of each experience. However, it supports the instantiation of experiences reconstructed or generated by the hippocampus during replay. The hippocampal network is made of complementary learning pathways: experiences are stored and retrieved in the CA3 autoassociative network, whereas novel experiences are generated from CA1. This division of labor within the hippocampus allows it to deal with the trade-off between memorization and generalization.

Key words

Generative model, Memory, Learning, Predictive coding, Neural network

Publications

• Fontaine, L., & Alexandre, F. (2025, June). Semantic and episodic memories in a predictive coding model of the neocortex. In 2025 International Joint Conference on Neural Networks (IJCNN) (pp. 1-9). IEEE.
• Fontaine, L., & Alexandre, F. (2026, September). A predictive coding model for memory replay in the neocortex. In International Conference on Artificial Neural Networks (pp. 15-26). Cham: Springer Nature Switzerland.

Jury

• M. Sander BOHTE, Professeur, Université d’Amsterdam, Rapporteur
• M. Alexandre PITTI, Professeur, Université de Cergy Pontoise, Rapporteur
• Mme. Hélène SAUZÉON, Professeure, Université de Bordeaux, Examinatrice
• M. Frédéric ALEXANDRE, Directeur de recherche, Université de Bordeaux, Directeur