Interview with C. Avare: “Thales needs smart chip / low power consumption. We are looking at Neural Network Chips as a possible solution”

Friday, September 4, 2015

Some weeks ago our partners Gina Alioto and Osman Unsal, from Barcelona Supercomputing Center, interviewed Christophe Avare, SOA and Big Data Segment Manager at Thales Research and Technology, the R&D entity of the French multinational company that helps its customers to make the right decisions at the right time and act accordingly in all its markets: aerospace, space, ground transportation, defence and security. Thales has operations in more than 50 countries; it has more than 68.000 employees and generated €13.03 billion in revenues in 2011. The Group is the 10th largest defence contractor in the world and a European technology leader in cryptographic and security architectures.

As a data processing specialist, Thales sees Big Data as a natural development of its core business and, accordingly, has invested consistently in research and development to write the complex algorithms needed to process that data, in addition to high-security storage solutions scaled to accommodate huge datasets.

Besides being a user of the common Big Data stacks and a Cloud provider, there are specific challenges in the area of the Hardware design for  embedded systems. Christophe Avare explains that Thales is “anticipating some major changes due to the influence of the Big Data approach, more specifically the resurgence of what is now called ‘Deep Learning’”. And he explains us that, “as an example, we are concerned by the so-called ‘Energy Efficiency Wall’ as this will not allow us to embed the required future processing capacity on autonomous systems, with a constraint power and thermal budget. As an example, we have started to closely look at the combination of artificial networks on chips and memristor technologies, because we think this is a promising way to go beyond this wall. But this is a long term R&D program”. The 20% of Thales Group revenues come from R&D. As it is its ability to understand very complex domains and deliver high-end systems, Thales needs a very active R&D and an efficient transfer into products.

As Christophe Avare explains to RETHINK big, Thales has many opportunities to deliver new functionalities in its products by providing better insight into not yet exploited data sources. More important, “the main source of opportunities for processing large data sets is generally in the most demanding environments”, says M. Avare.

And he exemplifies in “Autonomous Vehicles (like military drones), that have tight power budget (battery powered). For this business, this is a challenge, because the next generations of vehicles will have more sensors onboard, will support more complex and longer missions (and will even have to work in cooperation in a swarm-like behavior in the case of micro-drones). This will require more adaptation capabilities, a deeper situation understanding and so on, but the same power budget as the underlying technology is not evolving fast enough. Because they must be able to perform in situ processing - analyzing large inputs of data literally on the fly, the trend is to embed more deep learning or equivalent algorithms”.

M. Avare states that “Thales needs smart chip / low power consumption. We are looking at Neural Network Chips as a possible solution, and we are supporting academic research like MHANN (Memristive Hardware Artificial Networks)”. These Neural Networks “are the major way to have orders of magnitude in performance with a low power budget”.

In more speculative studies, Thales is interested by coupling memristors with such ANN chips. This will require a process able to provide billions of memristors on the same chip, but is the potential source of a major breakthrough.

Talking about emerging technology solutions, Christophe Avare explains that “Thales already uses Big Data and Big analytics platforms, but is most interested in unsupervised learning because the kind of data sets we need to use are usually not labeled. This is typically the case when one wants to have a chance to detect new cyber-attack patterns: nobody has seen them by definition! A real problem with supervised learning it that it is very difficult to collect and label large training sets, say for ANN, in a given domain (for example to perform adequate face recognition). The main problem with this approach is that whatever the kind of ANN you use, the first percents of accuracy are quickly achieved, but any further increase will require orders of magnitude more samples”.

Thales believes that this lack of training sets suggests business opportunities for:

  1. the creation and sales of high quality training sets
  2. a business in which I train the network with a specific purpose in mind and then sell it to you
  3. a shared directory of Data for machine learning without copyright


This interview is one of the several that the partners of RETHINK big are doing to the main stakeholders in order to identify the industry coordination points that will maximize European competitiveness in the processing and analysis of Big Data over the next 10 years.

Further information:

Other interviews: