Venue : Centre Broca Nouvelle-Aquitaine
On Zoom :
https://u-bordeaux-fr.zoom.us/j/83709626549
ID de réunion: 837 0962 6549
Thesis defended in english
Team: Computational neurosciences
Thesis directed by Frédéric Alexandre
Title
CONSEQUENCE: A Computational Model of the Interactions between Episodic Memory and Cognitive Control
Résumé
Episodic memory is often illustrated with the madeleine de Proust excerpt as the ability to re-experience a situation from the past following the perception of a stimulus. This simplistic scenario should not lead into thinking that memory works in isolation from other cognitive functions. On the contrary, memory operations treat highly processed information and are themselves modulated by executive functions in order to inform decision-making. This complex interplay can give rise to higher-level functions such as the ability to imagine potential future sequences of events by combining contextually relevant memories. How the brain implements this construction system is still largely a mystery. The objective of this thesis is to employ cognitive computational modeling methods to better understand the interactions between episodic memory, which is supported by the hippocampus, and cognitive control, which mainly involves the prefrontal cortex. It provides elements as to how episodic memory can help an agent to act. It is shown that neural episodic control, a fast and powerful method for reinforcement learning, is in fact mathematically close to the traditional Hopfield network, a model of associative memory that has greatly influenced the understanding of the hippocampus. Neural episodic control indeed fits within the universal Hopfield network framework, and it is demonstrated that it can be used to store and recall information, and that other kinds of Hopfield networks can be used for reinforcement learning. The question of how executive functions can control episodic memory operations is also tackled. A hippocampus-inspired network is constructed with as little assumption as possible and modulated with contextual information. The evaluation of performance according to the level at which contextual information is sent provides design principles for controlled episodic memory. Finally, a new biologically inspired model of one-shot sequence learning in the hippocampus is proposed. The model performs very well on multiple datasets while reproducing biological observations. It ascribes a new role to the recurrent collaterals of area CA3 and the asymmetric expansion of place fields, that is to disambiguate overlapping sequences by making retrospective splitter cells emerge. Implications for theories of the hippocampus are discussed and novel experimental predictions are derived.
Key words
Episodic memory, cognitive control, hippocampus, prefrontal cortex, decision-making, planning, imagination, computational neuroscience
Publications
- https://hal.science/hal-03885715
- https://inria.hal.science/hal-03359384
- https://inria.hal.science/hal-03359407
Jury
- Rapporteurs : Randall O’Reilly & Mehdi Khamassi
- Examinatrices et examinateurs : Anna Schapiro, Rufin VanRullen et Emmanuelle Abisset-Chavanne
- Directeur de thèse : Frédéric Alexandre