Numalis is an SME expert in software validation that markets design support tools. It was created in 2015 by two researchers: Arnault Ioualalen, Matthieu Martel, and an engineer: Nicolas Normand. It focuses on the industrialization of the research work conducted originaly by Arnault Ioualalen and Matthieu Martel. Based on its initial know-how Numalis has progressively developed the first neural network formal validation tools.
Numalis relies on more than 20 years of research and development on formal methods. Its history enables it to master a combination of know-how combining static code analysis through abstract interpretation, code synthesis and statistical analysis. Numalis is thus able to meet the challenges of code validation in several highly critical contexts (aeronautics, aerospace, defence, etc.).
Since 2015 Numalis provides tools for validating and improving the reliability of critical calculation code. The motivation for this work dates back to the series of industrial accidents that occurred during the 1990s (Patriot missile, Sleipner A platform, Ariane 5 rocket, etc.). Faced with these accidents the industry turned to research in the 90's to find ways to prevent or even correct these defects. Numalis then made this work available to help the industry prevent these disasters from happening again.
In 2017 Numalis starts using its existing know-how to extend its solutions to the validation of neural networks. Sponsored by the French Ministry of Defense (DGA), Numalis starts to answer the industrial imperative that has been expressed more recently on the validation of AI in critical systems. Indeed, the increasing use of deep-learning technologies in the industry raises several challenges. Although very efficient, deep-learning does not have sufficiently established validation methodologies and tools. Their widespread adoption is therefore slowed down in all critical or highly regulated environments (aeronautics, automotive, health, defence, etc.). As formal methods are one the strictest forms of validation that can be expressed on software, it is natural to see if they can allow to check the safety of such systems also. Numalis as an expert company in formal methods can help adapt its knowhow and toolset to accomplish such a new way of validating AI. This evolution is also powered by a normative vision which started to take shape in 2018 when the ISO subcommittee on AI was founded at its first plenary in Beijing. Today, Numalis is participating, along with the expert community, to the ISO 24029 series of standards. These documents will help engineers from all organizations develop more robust neural networks.
