Soutenance de thèse de Hugo Ladret

jeudi 8 février 2024 à 16:00
Publié le 25/01/2024

Jeudi 8 Février à 16h, Soutenance hybride hors-AMU. Salle 580-31, École d’Optométrie de l’Université de Montréal.

Lien Zoom https://umontreal.zoom.us/j/84293174953?pwd=YWlaZkREWURLTGw3S1BjQktFOHIwdz09

Hugo Ladret (Equipe NeOpTo)

Soutiendra sa thèse de doctorat intitulée :

A multiscale model to account for orientation selectivity in natural images

Présidente : Elvire Vaucher, Université de Montréal

Représentant du doyen : Matthieu Vanni, Université de Montréal

Rapporteuse : Laura Dugué, Université Paris Cité
Rapporteur : Nicholas Priebe, University of Texas in Austin

Examinateur : Olivier Marre, Institut de la Vision

Examinateur : Jason Eshraghian, University of California Santa Cruz

Examinateur : Lyle Muller, University of Western Ontario

Membre invité : Ede Rancz, Institut de Neurobiologie de la Méditerranée

Directeur de thèse : Laurent Perrinet, Institut des Neurosciences de la Timone
Codirecteur de thèse : Christian Casanova, Université de Montréal

Anyone who has ever tried to cross a street in Marseille knows the crucial role that visual uncertainty plays in our daily lives. From trying to read the ambiguous gaze of a driver on the Boulevard Baille to guessing whether a lion is hiding behind the grass in the savanna, humans implicitly factor sensory uncertainty into every decision they make. Since the days of Helmholtzian inference, a vast amount of psychophysical research has been devoted to this phenomenon, but its mapping to a specific neuronal circuit remains poorly understood.

This thesis sought to understand such neural bases, using a dual biological/computational approach. Based on the notion that uncertainty is directly tied to the (inverse) variance of sensory features, we will focus on manipulating low-level orientation content found in our everyday visual scenery, and seek to understand how this affects the primary visual cortex. We begin by exploring the computational advantages of using sparse, uncertain-encoding activity in neural networks to build representations of these natural images. This dictates functional principles that I  mapped to extracellular electrophysiology activity found across rodents, felines and macaques. In turns, this unravels new archetypal responses in the primary visual cortex that can be reproduced by models of canonical cortical recurrent circuitry. Based on the dynamics of these responses, I then propose that the principles guiding such uncertainty computations can be extended to extrastriate visual areas and even subcortical structures.

All in all, these four and a half years of research hammer on the same nail: cortical networks can compute the uncertainty of their activity based on recurrent synapses. This serves as a theoretical and experimental argument in favor of predictive coding under the free energy principle occurring in the cortex, paving the way to extend this research to other sensory modalities. Under this interpretation, uncertainty weighting dictates which of its own internal predictions versus its new external inputs the brain should trust to avoid getting run over in Marseille (or eaten in the savanna). This promises new avenues of research to understand how the brain updates and maintains a coherent representation of its environment.