Sites Inria

Version française

Neural networks

Estelle Raffin - 13/07/2016

Taking inspiration from the brain to design new memory systems

The Inria In’Tech seminar held on 23 and 24 June 2016, in association with NeuroSTIC Days, brought together researchers from a broad range of disciplines to share ideas around a common theme: artificial neural networks. Vincent Gripon, a researcher at Télécom Bretagne, explained his work which, inspired by the brain, entails designing new memory systems known as associative memories. He has also been involved in jointly organising the NeuroSTIC Days with Benoit Miramond and Martial Mermillod.

What applications do neural networks have in your research?

Vincent Gripon : Our approach is original because we are looking at memory disorders in neural networks, whereas most of our colleagues and the industry, in France and even in Europe, are focusing on learning and computation. In the brain, the long-term memory retains bits of information spanning an entire lifetime. However, during that time, the brain loses a significant proportion of the neurones and connections that it started out with, which we might call a loss of reliability. Now, with the increasingly dense stack of transistors on an electronics circuit board, we will see, for technical reasons, component malfunction and, therefore, a similar loss of reliability. How does an algorithm take account of such a loss of reliability? Here we come up against the same problem that we see in the brain. How come, even though, at the age of 60 or 70 years old we have lost some of our neurones, how come we are capable of remembering things from our childhood in vivid detail? My research starts from the point of trying to understand how the brain bypasses such a loss of reliability to try, by analogy, and propose systems that can efficiently remember information even if the architecture is unreliable.

Is this interdisciplinary approach important to you?

Vincent Gripon : All the scientists I've met agree that the future of research lies in an interdisciplinary approach. We really need different communities to meet and share ideas, because that's where you find inspiration. There is a huge difference between what we call human intelligence and what we call machine intelligence. To multiply two numbers each of which is thousands of digits long, you and I would require a certain amount of rigour and concentration. But a computer can do it in just a fraction of a second. Inversely, a child can identify a pizza in a photo instantaneously. But if we ask an algorithm or a computer to do the same thing, it's very complicated. The computers we work on do not function in anything like the same way as the brain. The kind of research that excites me is neuro-inspired: to look at how the only model we have of intelligence, the human brain, works, and see if there are any good ideas that we can use in this field to try and design neuro-inspired computers.

How did you become involved in organising NeuroSTIC Days?

Vincent Gripon : My team and I have attended the NeuroSTIC Days for many years. They bring together many research networks (GDR research groups) and many different disciplines between which there is often very little communication. For me, NeuroSTIC Days are the perfect manifestation of the desire for dialogue between disciplines, in a pretty relaxed atmosphere, and to encourage academic and industrial objectives to converge. For instance, during the days, I talked with some psychologists for whom connecting with industry is not so obvious. However, we sometimes find that there really are things that we can work on together. I agreed to get involved in organising NeuroSTIC Days because I believe it is our duty to encourage this kind of initiative and inspire collaborative work. It's a great opportunity to get a message across to young students and PhD students, namely, that the future means having the courage and the desire to go out and meet with psychologists, doctors, electronics engineers, computer scientists, signal processing experts, etc., and to realise just how stimulating the crossover of ideas can be for research.

NeuroSTIC Days and In’Tech 2016 seminar slideshow

  • Probayes demonstration - Deep learning: embed more neurones in your applications © Inria / Photo E. Raffin

  • Demonstration of machine-learning using GPUs by François Courteille from NVID © Inria / Photo E. Raffin

  • Demonstration of PNeuro, the Neural Network accelerator © Inria / Photo E. Raffin

  • Nearly 200 people came to attend the seminar © Inria / Photo E. Raffin

  • Olivier Brousse from Global Sensing Technologies presents a low power accelerator © Inria / Photo E. Raffin

  • Rodolphe Heliot from Schneider Electric talks about requirements in hardware and algorithms © Inria / Photo E.Raffin

  • Georges Malliaras, from Mines St-Etienne, presents organic electronics © Inria / Photo E.Raffin

  • In’Tech seminars and NeuroSTIC Days are excellent occasions for networking and sharing ideas © Inria / Photo E.Raffin

Keywords: Vincent Gripon IN'Tech INRIA Grenoble - Rhône-Alpes Associative memories Neuro-inspired NeuroSTIC Days