Laboratoire Matière et Systèmes Complexes

Thèse de Renaud Renault effectuée sous la direction de Samuel Bottani et Pascal Monceau avec la collaboration de Elisha Moses (Institut Weizmann) et de Jean-Louis Viovy (Institut Curie).

Mémoire de thèse

Conception émergente de processeurs neuronaux

Soutenance le lundi 12 Octobre 2015 à 14h.

Lieu : bâtiment Halle aux Farines, Salles des Thèses (580A, 5e étage), 2 rue Marguerite Duras 75013 PARIS

La soutenance sera suivie d’un pot au sixième étage du bâtiment Condorcet, 10 rue A. Domon et L. Duquet, en face.

Abstract  :

When seeded on an appropriate substrate, neurons extracted from brains are able to reconnect and communicate with each other. Depending on how they connect, they generate different patterns of electrical activity that can be exploited to create, store, or process information. Such neuronal devices are amenable to many ground-breaking applications, including glucose powered neural implants to repair cerebral functions, in-vitro models for cognitive studies, and even new kinds of artificial intelligence. However, there is currently no consensus about how to harness the processing capabilities of neurons in culture or how to build robust and efficient neuronal devices. My work therefore focuses on new designs, inspired by Rosenblatt’s perceptron, that can bypass the limitations of previous neuronal devices. In these architectures, the neuronal devices are divided into several neuronal populations (individually acting as on-off « switch-like » units), by the mean of compartmented microchips.

Between the compartments, axon guides of various geometries were used to precisely control the directions and strengths of the connections between units, thus defining the computational abilities of the devices. The capacity of perceptrons to learn new computations by adjusting the strength of the connections between their units can also be implemented in such neuronal devices by exploiting synaptic plasticity, the natural ability of living neurons to adapt the strength of their connections when subjected to specific stimulations.

During my defense, I will present a model, known as quorum percolation, which links neuronal populations « switching » behaviors with experimentally accessible parameters such as their density, size and composition. Another model based on experimental data was developed to predict axonal growth under topographical constraints, completing the toolbox for the in-silico study and design of neuronal devices. In parallel, a fully optical interface, based on optogenetics and calcium imaging to stimulate and read neuronal activity respectively, allowed quantitative measurements of functional connectivity and synaptic plasticity inside neuronal devices. This interface was furthermore implemented using low-cost off-the-shelf components, making it available to virtually any lab. All these theoretical and technological development laid a solid fondation for the systematic study of communication, computation, plasticity and learning inside living neuronal devices, and their use in future practical applications.

Keywords :

Neuronal cultures, neurophysics, neuro-physics, percolation, neuronal engineering, neuronal growth, synaptic plasticity.

References :

R. Renault, P. Monceau and S. Bottani, Phys. Rev. E. 88, 062134 (Dec. 2013). Memory decay and loss of criticality in quorum percolation.

R. Renault, P. Monceau, S. Bottani and S. Métens, Physica A. 414, 352, (Juil. 2014). Effective non-universality of the quorum percolation model on directed graphs with Gaussian in-degree.

Renaud Renault, Nirit Sukenik, Stéphanie Descroix, Laurent Malaquin, Jean-Louis Viovy, Jean-Michel Peyrin, Samuel Bottani, Pascal Monceau, Elisha Moses, Maéva Vignes, PLoS ONE 10(4) : e0120680. doi:10.1371/journal.pone.0120680, (Apr. 2015). Combining microfluidics, optogenetics and calcium imaging to study neuronal communication in vitro.