Description du projet
Des systèmes d’IA portables et plus efficaces qui fonctionnent comme des cerveaux humains
Les algorithmes actuels d’IA les plus performants s’inspirent de réseaux neuronaux qui s’apparentent à ceux du cerveau. Cela dit, contrairement à nos cerveaux d’une grande efficacité, l’exécution de ces algorithmes sur des ordinateurs consomme de très grandes quantités d’énergie. Ces unités centrales de traitement extrêmement inefficaces entravent le développement de systèmes d’IA efficaces, évolutifs et portables. Le projet ChipAI, financé par l’UE, exploitera le potentiel des nanotechnologies photoniques pour mettre au point des unités centrales de traitement compactes, à large bande passante et efficaces sur le plan énergétique. Les chercheurs utiliseront des nanostructures à semi-conducteurs à effet tunnel résonnant intégrées dans des cavités métalliques de sous-longueurs d’ondes 100 fois plus petites que les cavités conventionnelles afin de confiner, d’émettre et de détecter efficacement la lumière. Les résultats du projet ouvriront la voie à de nouveaux développements dans le domaine balbutiant de l’informatique optique neuromorphique.
Objectif
The same way the internet revolutionized our society, the rise of Artificial Intelligence (AI) that can learn without the need of explicit instructions is transforming our life. AI uses brain inspired neural network algorithms powered by computers. However, these central processing units (CPU) are extremely energy inefficient at implementing these tasks. This represents a major bottleneck for energy efficient, scalable and portable AI systems. Reducing the energy consumption of the massively dense interconnects in existing CPUs needed to emulate complex brain functions is a major challenge. ChipAI aims at developing a nanoscale photonics-enabled technology capable of deliver compact, high-bandwidth and energy efficiency CPUs using optically interconnected spiking neuron-like sources and detectors. ChipAI will pursue its main goal through the exploitation of Resonant Tunnelling (RT) semiconductor nanostructures embedded in sub-wavelength metal cavities, with dimensions 100 times smaller over conventional devices, for efficient light confinement, emission and detection. Key elements developed are non-linear RT nanoscale lasers, LEDs, detectors, and synaptic optical links on silicon substrates to make an economically viable technology. This platform will be able to fire and detect neuron-like light-spiking (pulsed) signals at rates 1 billion times faster than biological neurons (>10 GHz per spike rates) and requiring ultralow energy (<10 fJ). This radically new architecture will be tested for spike-encoding information processing towards validation for use in artificial neural networks. This will enable the development of real-time and offline portable AI and neuromorphic (brain-like) CPUs. In perspective, ChipAI will not only lay the foundations of the new field of neuromorphic optical computing, as will enable new non-AI functional applications in biosensing, imaging and many other fields where masses of cheap miniaturized pulsed sources and detectors are needed.
Champ scientifique
- engineering and technologyelectrical engineering, electronic engineering, information engineeringelectronic engineeringcomputer hardwarecomputer processors
- natural sciencesbiological sciencesneurobiologycomputational neuroscience
- natural sciencescomputer and information sciencesdata sciencedata processing
- natural sciencescomputer and information sciencesartificial intelligencecomputational intelligence
Mots‑clés
Programme(s)
Régime de financement
RIA - Research and Innovation actionCoordinateur
4715-330 Braga
Portugal