You are here : Home > More efficient neural network execution

News | New technologies | Embedded systems

Embedded systems

More efficient neural network execution

​List, a CEA Tech institute, is speeding up neural network execution. The networks—inspired by biological neurons—are driving a new generation of products at GlobalSensing Technologies, a Dijon, France-based startup.

Published on 2 February 2016

Neural networks are playing an increasing role in technologies to recognize images, sounds, or any other signal generated by a sensor or network of sensors—giving the devices the ability to interpret what they perceive. GlobalSensing Technologies uses neural networks in the artificial intelligence inside its embedded products. With applications in industries like food, video surveillance, and manufacturing, there are virtually no limits to the uses of neural networks, which help "recognize" specific environments or situations and trigger an appropriate response.

GlobalSensing Technologies has set up a joint lab with List that has developed a processor whose architecture speeds up neural network execution. "The calculations required for this type of technology are simple, but great in number. The architecture we developed has the smallest, most efficient calculation units possible," said a List representative. How the data gets to the calculation units is also vital. The lab tested the architecture on a field-programmable logic gate array, which enables the rapid prototyping of a complete system.

The results are very promising, with 96% recognition rates—excellent for a very simple neural network—and a 60% improvement in energy efficiency over competing solutions if the application is ported to a dedicated FSOI* component.

Several patents have been obtained for the new architecture, giving GlobalSensing Technologies a crucial competitive advantage on a high-growth market.


*Fully-depleted silicon-on-insulator: an innovative transistor technology with a thin insulating layer and a thin silicon film developed by Leti, a CEA Tech institute.

Top page

Top page