On Monday September 04^th, at 14:00, Manon DAMPFHOFFER (SPINTEC) will defend her PhD thesis entitled :
Models and algorithms for implementing energy-efficient spiking neural networks on neuromorphic hardware at the edge

Place : GreEn-ER – Grenoble INP- Ense3, amphi 2A006

zoom link : https://grenoble-inp.zoom.us/j/3741741553
ID : 374 174 1553
Password : 062839

Abstract : Deep learning in Artificial Neural Networks (ANNs), a branch of artificial intelligence (AI), is considered a revolution in computing and is impacting every sectors of the economy. However, ANNs are very compute- and memory- intensive, which limits their integration into edge devices for embedded applications. Spiking Neural Networks (SNNs) are promising energy efficient alternatives to ANNs and hence are good candidates for edge AI implementations. However, the gap between the algorithmic development of SNNs on the one hand, and their hardware implementation on the other hand, makes it difficult to achieve truly efficient solutions. In this context, this thesis proposes models and algorithms to improve the accuracy and energy efficiency of SNNs from an algorithm-hardware co-development perspective. First, in the interests of comparing SNNs and ANNs implementations on dedicated neural network accelerators, a high-fidelity model of their energy efficiency is provided. In particular, it is found that spike sparsity plays a key role in the efficiency of SNNs. Second, based on the previous conclusions, a novel SNN model, SpikGRU, is proposed and its energy efficiency is demonstrated. Third, the use of analog Non-Volatile Memories is considered to implement the synaptic weights of ANNs and SNNs. With an adapted training methodology, ANNs and SNNs are demonstrated to be very robust to read errors due to the variability of analog devices. By promoting algorithm-hardware co-development, this work aims at paving the way for efficient neural network implementations at the edge.

jury :
Timothée Masquelier – Directeur de recherche, CNRS – Université de Toulouse 3, Rapporteur
Benoît Miramond – Professeur des Universités, Université Côte d’Azur, Rapporteur
Melika Payvand – Assistant Professor, Université de Zurich et ETH Zurich, Examinatrice
Pascal Perrier – Professeur des Universités, Université Grenoble Alpes / Grenoble INP, Examinateur
Damien Querlioz – Chercheur HDR, CNRS – Université Paris-Saclay, Examinateur

Thesis supervisors :
Lorena Anghel – Professeure des Universités, Université Grenoble Alpes, CEA, CNRS, Grenoble INP, INAC-Spintec, Directrice de thèse
Alexandre Valentian – Ingénieur Docteur, Chercheur, CEA, LIST, Co-encadrant de thèse, Invité
Thomas Mesquida – Ingénieur Docteur, Chercheur, CEA, LIST, Co-encadrant de thèse, Invité