Skip to main content
Aller à la page d’accueil de la Commission européenne (s’ouvre dans une nouvelle fenêtre)
français fr
CORDIS - Résultats de la recherche de l’UE
CORDIS

Energy-based Learning of Complex Dynamical Systems

Objectif

As digital computing approaches its limits, a paradigm shift is needed to push the boundaries of computation. Neuromorphic computing, an analog computing paradigm inspired by the human brain, offers a promising alternative. Implementing mathematical models such as artificial neural networks directly in neuromorphic hardware holds tremendous potential: it can drastically reduce energy consumption while enabling faster inference when compared to digital architectures. Yet, a major bottleneck is the lack of dedicated algorithms for training analog circuits from data. Traditional methods such as backpropagation are unsuitable, as they require different circuits for training and inference, thus introducing energy-costly data transfer between circuits.

I propose a novel systems theory for learning in analog circuits, inspired by ideas from the literature on energy-based models, pioneered by Hopfield, Hinton and others. My approach is to cast the learning problem as the feedback interconnection of continuous-time circuit dynamics with an optimization algorithm. Using this framework, I will develop novel energy-based learning algorithms, capable of training circuits from input-output data. The developed algorithms are fully distributed, thus enabling efficient training of large-scale circuits, as well as robust to uncertain circuit elements, therefore accounting for device-to-device variations in neuromorphic hardware.

My preliminary results on resistive circuits already demonstrate the potential of this approach. Building on this foundation, the project will establish a comprehensive system-theoretic framework, capable of handling nonlinear and dynamic circuits with (mem)ristive and capacitive elements, as well as more general dissipative networks. This fundamental research project has the potential to significantly advance the theory of physics-based learning. The proposed theory will be supported by efficient computational tools and tested in neuromorphic hardware.

Champ scientifique (EuroSciVoc)

Mots‑clés

Programme(s)

Thème(s)

Régime de financement

HORIZON-ERC - HORIZON ERC Grants

Voir tous les projets financés dans le cadre de ce programme de financement

Appel à propositions

(s’ouvre dans une nouvelle fenêtre) ERC-2026-STG

Voir tous les projets financés au titre de cet appel

Institution d’accueil

RIJKSUNIVERSITEIT GRONINGEN
Contribution nette de l'UE
€ 1 499 620,00
Coût total
€ 1 499 620,00

Bénéficiaires (1)