NeurEco: The Parsimonious Neural Network tools that reduces the CARBON IMPACT of AI       by several orders of magnitude.

 

Impacts of parsimony:
  • Deep learning of continuous phenomena - Unlike the state of the art, which is oriented to discrete responses, parsimonious neural networks are suitable for both continuous and discrete responses and are capable of capturing the physical or biological phenomena conveyed by the data.
  • Ability to create reliable complex dynamic models - This can be achieved even in the case of quasi-chaotic phenomena, where parsimony is of critical importance and any redundancy would have irremediable consequences on the quality of the model.
  • Robustness – our neural networks are capable of resisting "deepfool" attacks, which can provide a significant challenge for the redundant state-of-the-art models.
  • Embedding of artificial intelligence algorithms  to meet the challenges of real-time control of complex systems("Digital Twin");
  • Processing of hybrid data (simulation / measurements).
Our parsimonious neural networks represent a major step forward from the state of the art.

A major industrial player put ADAGOS in competition with one of the world-leaders in AI on the problem of predicting turbocharger blade deflections. Without a doubt, the results were in our favor:

The world-leader’s network

ADAGOS’ parsimonious network

 

Leading Tech Compagny ADAGOS
Learning data
(Percentage of available data)
1 000 000 samples
(100%)
10 000 samples
(1%)
Size of the network (Number of links) 500 000 194
Computing resources A 15 000 $ GPU card 1 minute 13'' on a laptop
Energy consumption 1 kWh 0.0005 kWh

NeurEco Framework

deepROM®

deepROM® is our neural network factory. It excels at figuring out automatically the optimal network topology. The whole process is then standalone and avoids overlearning. Hence the user only has to provide the learning data and shortly after gets a reliable model. See examples here.

xROM™

xROM™ is our implementation of recurrent reduced order models. It builds models for nonlinear dynamic processes and get rid of their inherent instabilities. See examples here.

coROM™

Our convolutional neural network solution is adapted to higher dimensional problems. It gets rid of the so-called curse of dimensionality. It controls adaptively the data generation process. See examples here.

Our main applications

Our partners

What our clients say

En rupture totale par rapport à l’état de l’art, la société ADAGOS arrive à créer des modèles réduits ayant un grand nombre de paramètres (plusieurs centaines dans les cas considérés). Cela veut dire que ces outils ne semblent pas être affectés par ce que l’on appelle la malédiction de la dimension (curse of dimensionality). Cela parait d’autant plus étonnant que nos paramètres varient dans des intervalles très larges et ont un caractère discret aussi bien qualitatif que quantitatif.

Jean-Marie Hamy Technical Project Manager and Design Authority @ Framatome

Les enjeux sont considérables pour les systèmes embarqués et les applications médicales. De ce fait, nous maintenons une veille technologique et nous sommes en contact avec les universités les plus prestigieuses pour scruter la moindre évolution dans le domaine. Mais, il n’était pas nécessaire d’aller chercher bien loin, l’innovation de rupture peut venir d’une petite startup toulousaine appelée ADAGOS.

Michel Rochette Director of research @ Ansys France