You are a few to blame (kindly) for this blog a certain orientation towards American research. It is not a will, but a consequence (1) of the consequent US budgets in terms of defense R & D and (2) a certain control of communication by our neighbors on the other side of the Atlantic To communicate with a certain talent of the political programs having little to do with the reality, I say that way ...).
Once a year, the DGA organizes its Innovation Forum. It was yesterday and today, on the site of Palaiseau, and at the time, it brings a real active will of communication on the part of both institutions, but also laboratories and small businesses. So I made my innovation market, here is a first result.
The project is called AXONE. It is an artificial neural system capable of performing real-time scene analysis tasks. This is the result of a RAPID project (the Support Scheme for Duale Innovation - see the reference page here ) involving the Saint-Louis Institute, Spikenet Technologies and GlobalSensing Technologies.
The idea is to use a certain type of artificial neural network, Spike neurons, to perform real-time analysis of a visual scene and to embed them on dedicated components (SoC or Systems on Chips ). I will try to explain the concept simply - especially since I worked more than twenty years ago with Professor Simon Thorpe, creator of SpikeNet technology (it seems to me that I even committed An article on the subject ...).
To make it simple: in neurobiology, there has long been the question of the nature of the coding of information by neurons. The general theory is based on a frequential coding of the information (frequency of electric discharges). But there is another theory based on a temporal coding of the information: the coding would be done by pulses (spikes) or more precisely by the times of emission of the pulses.
The temporal aspect is thus taken into account. An artificial network of spike neurons is designed to simulate networks that contain a very large number of asynchronous discharge neurons and learn by coding discharge sequences. This is called row encoding (obviously, I simplify). This technology is widely used for the recognition of shapes, and in particular image processing.
The advantage of this technology is that the learning time is very fast, and very tolerant (under illumination conditions, noise, contrasts ...). In the AXONE project, participants were able to implement a spike neural network on a dedicated board (called an FPGA processor).
Basically, it is a reconfigurable processor, comprising 1024 artificial neurons, and designed by GlobalSensing Technologies. With SpikeNet and ISL, and in 24 months, AXONE project actors realized a reconfigurable camera generating Spikes in place of the images. The work consisted of (besides obviously the underlying algorithmic) to integrate this artificial neural network with a sensor, within a stand-alone camera, and to develop the software library for the implementation of these components.
And the result is impressive. During the presentation at the DGA Forum, we were able to visualize the recognition of faces in real time (each face is recognized, in real time, with its unique signature). There are many applications: security and surveillance of sensitive sites, with the camera hiding itself, abandoned sensors capable of performing an in situ analysis (see my article on on-board AI) , and of course increasing the capacity of Recognition of shape in robotics and in particular for drones.
I would add that the DGA has taken an original initiative: that of sponsoring certain innovations by personalities of the institution. In this case, AXONE is sponsored by the excellent Lionel MORIN, director of the CATOD (Technological-Operational Defense Analysis Center) - below.
A technology to follow, and an excellent illustration of the innovation capabilities of the French defense ecosystem - I will soon publish other articles following my visit to the Innovation forum.
BRN Price at posting:
27.0¢ Sentiment: Buy Disclosure: Held
(20min delay)