Skip to Main content Skip to Navigation

Conception de réseaux de neurones sur silicium à l’aide de synapses memristives : application au traitement d’image

Abstract : Nowadays, the artificial intelligence is a technology used more and more in diverse features. The usage of their algorithms need a huge consummation of energy, for environmental and societal issues it is necessary to reduce their power consumption. The classical von Neumann architecture used in traditional computer is not efficient for artificial intelligence algorithms in terms of energy and calculation speed. The European project ULPEC which this thesis takes part, has in aim to design an ultra-low power bio-inspired neural network based on memristive synapses and an event-based camera. Memristors which play the role as synapses, are variable resistors controllable by voltage at their terminals.The purpose of this project is to make a chip embedding an event camera, a matrix of 784x100 memristors and design the analog neural network to achieve the best learning as possible. The neural network is composed of 784 input neurons and 100 output neurons, where each input neuron is directly connected to a single one pixel from the camera. Output neurons make the images recognition and modify the weight of their synapses to realize the learning.Neurons are modelized by a LIF neuron (Leaky Integrated and Fire), this model is close to the biology and has the advantage to be easy to design and implemented by electronics. We use a current conveyor into the output neuron to implement this model. We have built a simulator close to the physics of the neural network, and we have obtained 67% of recognition on handwritten figures database.This thesis has aimed to explore the feasibility of the neural network base on synapses memristive realization to highlight the issue and solve it with the technical solution when it is possible. These works allowed to explore the feasibility of this project and give clue for future projects.
Document type :
Complete list of metadata
Contributor : ABES STAR :  Contact
Submitted on : Friday, February 4, 2022 - 9:15:08 AM
Last modification on : Saturday, February 5, 2022 - 3:38:24 AM
Long-term archiving on: : Thursday, May 5, 2022 - 6:27:57 PM


Version validated by the jury (STAR)


  • HAL Id : tel-03556410, version 1


Charly Meyer. Conception de réseaux de neurones sur silicium à l’aide de synapses memristives : application au traitement d’image. Electronique. Université de Bordeaux, 2021. Français. ⟨NNT : 2021BORD0189⟩. ⟨tel-03556410⟩



Record views


Files downloads