Description du projet
De nouvelles puces neuromorphiques pour faciliter l’apprentissage sur plusieurs échelles de temps
L’informatique neuromorphique est un terme générique qui désigne un ensemble d’efforts visant à développer des systèmes de calcul imitant la structure neurale du cerveau humain. Le projet MeM-Scales, financé par l’UE, travaillera au développement d’une plateforme innovante qui servira de base à de futurs produits combinant une efficacité énergétique extrême et des capacités cognitives humaines. L’accent sera placé sur la mise au point de nouvelles technologies de mémoire et de dispositifs, ainsi que sur des algorithmes d’apprentissage autonome qui prennent en charge l’apprentissage sur puce sur plusieurs échelles de temps, tant pour les synapses que pour les neurones. Ce nouveau type de technologie informatique pourrait être utilisé pour faire progresser la surveillance distribuée de l’environnement, les micropuces de diagnostic médical implantables, l’électronique portable et l’interaction homme-machine.
Objectif
"The project MeM-Scales aims at lifting neuromorphic computing in analog spiking microprocessors to an entirely new level of performance. Work in this project is based on a dedicated commitment that novel hardware and novel computational concepts must be co-evolved in a close interaction between nano-electronic device engineering, circuit and microprocessor design, fabrication technology and computing science (machine learning and nonlinear modeling). A key to reflecting ""hardware physics"" in ""computational function"" and vice versa is the fundamental role played by multiple timescales. Here MeM-Scales introduces a number of innovations. On the side of physical substrates, novel memory and device technologies, supporting on-chip learning over multiple timescales for both synapses and neurons, will be fabricated. To enable timescales spanning up to 9 (!) orders of magnitude both volatile memory and non-volatile memory as well as Thin Film Transistor technology will be exploited. On the side of computational theory, autonomous learning algorithms and architectures supporting computation over these wide range of timescales will be developed. These computational methods are specifically tailored to cope with the low numerical precision, parameter drift, stochasticity, and device mismatch which are inherent in analog nano-scale devices. These cross-disciplinary efforts will lead to the fabrication of an innovative hardware/software platform as a basis for future products which combine extreme power efficiency with robust cognitive computing capabilities. This new kind of computing technology will open new perspectives, for instance, for high-dimensional distributed environmental monitoring, implantable medical diagnostic microchips, wearable electronics or human-computer interfacing."
Champ scientifique
- natural sciencescomputer and information sciencessoftware
- engineering and technologyelectrical engineering, electronic engineering, information engineeringelectronic engineeringcomputer hardwarecomputer processors
- natural sciencescomputer and information sciencescomputational science
- engineering and technologymaterials engineeringcoating and films
- natural sciencescomputer and information sciencesartificial intelligencemachine learning
Mots‑clés
Programme(s)
Régime de financement
RIA - Research and Innovation actionCoordinateur
75015 PARIS 15
France