Description du projet
Faire progresser la technologie pour un calcul efficace inspiré par le cerveau
L’informatique neuromorphique est une branche de l’informatique et de l’ingénierie qui s’inspire de l’architecture et du fonctionnement du cerveau humain pour concevoir et construire des systèmes informatiques plus efficaces et plus proches du cerveau. Les memristors ont la capacité d’émuler les synapses dans les systèmes biologiques et ont donc suscité l’intérêt pour leur utilisation dans l’informatique neuromorphique. Financé par le Conseil européen de la recherche, le projet RobustNanoNet vise à résoudre les problèmes de performance des memristors en améliorant la technologie de commutation résistive. Les chercheurs développeront des dispositifs avec de nouveaux matériaux, testeront leurs performances et étudieront leur utilisation dans des systèmes informatiques avancés pour l’apprentissage automatique. L’objectif est de créer une technologie fiable pour un calcul efficace et fiable inspiré par le cerveau.
Objectif
Resistive switching refers to the controlled change in resistance of an electronic material, e.g. metal oxide, via the creation and modulation of nanoscale filaments. Although its physics is not yet fully understood, resistive switching devices (called memristors) are promising as efficient artificial synapses in neuro-inspired computing systems. However practical challenges exist. Current devices excel in only a few of the performance metrics necessary for circuit and system integration. Moreover, they exhibit non-idealities causing neuromorphic systems using these devices to have low performance. The project will address this key issue by pursuing device-system co-optimization across four objectives, aiming to engineer a single “hero” resistive switching technology with all the desired metrics. Aim 1 will develop resistive switching devices based on a new class of materials with broad compositional space, called high entropy oxides. Promising compositions will be fabricated in a high throughput fashion. In Aim 2, a proposed characterization method via a state-of-the-art mid-infrared laser will help understand in-operando the filamentary switching at nanoscale and uncover the physical mechanisms behind its non-idealities. The fabrication and characterization will iteratively target a broad range of performance metrics. Some metrics can only be quantified across a population of devices, so Aim 3 will integrate the optimized devices on transistor circuitry for benchmarking at scale. Aim 4 targets the applicability of these devices to next generation neuromorphic systems for machine learning training. Preliminary work on a multi-layer neural network validated this concept and indicated the need for co-optimization, as proposed. RobustNanoNet will address the interdisciplinary challenges towards a reliable resistive switching technology to support robust neuromorphic systems for energy efficient computing.
Champ scientifique
- natural scienceschemical sciencesinorganic chemistryinorganic compounds
- natural sciencescomputer and information sciencesartificial intelligencemachine learning
- natural sciencesphysical sciencesopticslaser physics
- natural sciencescomputer and information sciencesartificial intelligencecomputational intelligence
Mots‑clés
Programme(s)
- HORIZON.1.1 - European Research Council (ERC) Main Programme
Thème(s)
Régime de financement
HORIZON-ERC - HORIZON ERC GrantsInstitution d’accueil
077190 Voluntari
Roumanie