Neural networks for the representation of 3D environments
Paul UTEZA
Team CENTURI

Abstract
Animals form internal representations of their environment to perform complex tasks. For example, during spatial navigation, they carry out sophisticated operations to optimize their trajectories. Yet, understanding the brain activity underlying these computations remains an active and open area of research. Most investigations to date have been conducted in simplified 2D environments. While such studies have identified several environment-encoding cells (e.g., place cells, grid cells), it is still crucial to decipher how more complex 3D environments are encoded.

To address this, we propose leveraging deep learning techniques capable of forming the representations required for such environments. We examine how a recurrent network performs tasks analogous to mental rotation, analyze the architecture of such networks, and characterize their dynamics. Next, animals are observed performing navigation tasks within large-scale, semi-natural virtual settings with unpredictable rewards, reflecting the uncertainty and search dynamics that characterize natural exploration. We demonstrate that animals can infer hidden task structure, favor generalized schema-like representations rather than specific zones, and adapt their reward-seeking strategies over time.

Ultimately, this work highlights the value of integrating experimental observations with computational modeling to advance understanding of how agents —both biological and artificial— represent and manipulate space, refining current theories of spatial cognition and demonstrating the importance of bridging neuroscience and machine learning to uncover deeper insights into the richness of spatial cognition.

Jury
Albert COMPTE – FCRB-IDIBAPS, Barcelone – Rapporteur
Gianluigi MONGILLO – Institut de la Vision, Paris – Rapporteur
Anna MONTAIGNINI – Institut de Neurosciences de la Timone, Marseille – Examinatrice
Guillaume DUGUÉ – Institut de Biologie de l’École Normale Supérieure, Paris – Examinateur
Laurent RAYMOND – Centre de Physique Théorique, Marseille – Président du jury
Hervé ROUAULT – Centre de Physique Théorique, Marseille – Directeur de thèse

Wednesday 24th September at 2pm – Amphi 5 at the CPT – Centre de physique théorique, Luminy Campus