The early days of artificial intelligence in France
He told us how in 1986 he came to create the trail-blazing research laboratory, DIAM, where many renowned French researchers were trained. Prof. Azencott also instituted ENS Paris Saclay’s highly reputed MVA Master’s programme.
You were among the first to get into artificial intelligence in France, in the 1980s. What were the circumstances?
Robert Azencott: It was at Les Houches, in 1983. I was attending an IBM seminar on statistical physics. I was a professor at ENS Ulm back then. The seminar was organized by Marc Mézard and other notable physicists such as David Sherrington and Scott Kirkpatrick—the leading specialists at the time of simulated annealing, a stochastic optimization method related to interacting particle systems.
I was intrigued by several presentations on distributed memory systems in artificial neural networks, a concept that John Hopfield (Nobel Prize in Physics) had recently explored. My “conversion” to artificial neural networks and automatic image analysis was sealed a year later, when I read a particularly incisive article by Donald and Stuart Geman on artificial vision.
I had taught for several years in the United States and I was very impressed by the pace of innovation across the Atlantic. Over there, after the boom (and bust) of “expert systems” in 1965-1975, artificial neuron networks were gaining momentum. The Defense Advanced Research Projects Agency (DARPA) was investing heavily in this research that aimed to emulate human intelligence. It was being talked about in France, of course, but very little among mathematicians. I quickly became convinced that AI offered exciting research prospects for applied mathematics.
In France, were these subjects mainly of interest to physicists?
R.A.: Yes. At the time, it was physicists such as Marc Mézard and his colleagues at ENS Ulm, Christian Jutten and Jeanny Hérault in Grenoble, and neurobiologists such as Jean-Pierre Changeux at the Collège de France who were exploring the potential of neural networks—real or artificial ones.
In 1984, French mathematicians were only interested in AI from a distance, partly because proving fine theorems in these areas of application was a seemingly impossible challenge. As for applied computer science teams, such as Olivier Faugeras’s at INRIA and Henri Maitre’s at Telecom Paris; they were certainly remarkably active in artificial vision and automatic image analysis, but with very Cartesian algorithmic techniques, far from the empirical pragmatism of machine learning.
French computer scientists working on artificial vision had reliability and robustness constraints due to their industrial collaborations, so were initially reluctant to explore much less predictable techniques such as neural networks. Their interest in these approaches came later, in the early 1990s.
Whereas you didn’t hesitate to take the plunge and convert to artificial intelligence for your image analysis research...
R.A.: I got into this fascinating field in 1984-85, at first by devouring some 6,000 pages of American scientific publications on artificial vision. I was fascinated at the thought of understanding how human intelligence works, how speech and vision is learned, how millions of images are mentally represented, and the ability to learn to read, write, move, etc.
I was hooked, and determined to start intensive research in the field. So in 1986, I created a new laboratory: DIAM, an English acronym for Distributed Intelligence And Mathematics, which was soon renamed the Département d’Intelligence Artificielle et Mathématiques. It was not, strictly speaking, a department; I was the only professor, surrounded by a dozen PhD students. We had the major (and unreachable!) target of understanding how intelligence emerges when it is “distributed” over millions of decisional units that communicate with each other.
In practice, my goal was to explore two interrelated areas of mathematics research: Markov Random Fields for digital image analysis and machine learning through stochastic recurrent neural networks (Botzmann Machines).
Didn’t many French mathematicians begin their research in this laboratory?
R.A.: I managed to recruit a brilliant team of doctoral students in applied mathematics. Most of them now head university or industrial laboratories, in France and abroad. They include Laurent Younes, Olivier Catoni, Alain Trouvé, Jérôme Lacaille, Isabelle Gaudron, Mathilde Mougeot, Oussama Cherif, François Coldefy, Jia Ping Wang, Jen Feng Yao, Michel Lévy, Bernard Chalmond, Christine Graffigne, Charles-Albert Lehalle, Bruno Durand, Pierre Herbin, Eric Lhomer and many more.
DIAM was a research venture that lasted 10 years, until my wife’s sudden death in 1996. We developed and applied innovative AI techniques in many industrial contexts with EDF, CNES, SNECMA, ARIANE Espace, etc. DIAM researchers made advances in simulated annealing, Markov chain image analysis, synchronous Boltzmann machines, pattern and sound recognition, etc.
In 1993, when I left Paris XI to come to ENS Cachan, DIAM was incorporated into CMLA (Centre for Mathematics Studies and Applications) under the impetus of Jean-Michel Ghidaglia.
In this new setting in 1995, you started the MVA (Mathematics Vision Learning) Master’s programme, which remains a reference for artificial intelligence to this day. Was that another first?
R.A.: A first in the Paris region in any case (there were probably similar courses in Grenoble). I was able to bring together excellent specialists thanks to collaboration with ENS (Ulm and Cachan), Ecole Polytechnique, Telecom Paris, University Paris XI, INRIA, and more.
Unsurprisingly, top students were soon flocking to the programme. And thanks to the remarkable mathematicians who joined CMLA, including Yves Meyer, Jean-Michel Morel, Lionel Moisan, Alain Trouvé and Nicolas Vayatis, CMLA has become a dynamic hub for research and innovation in artificial vision and machine learning.
What computing resources did you have at the time?
R.A.: That is one of the key points I identified early on. In 1988, in collaboration with other labs, we managed to get the Ministry of Defence to finance the purchase of a massively parallel supercomputer, the Connection Machine (64,000 processors), the first and only one in France at the time. Others followed, including a MASPAR (4,096 processors).
Using these supercomputers, Jérôme Lacaille, Laurent Younes, Oussama Cherif and I were able to implement artificial retinas based on Botzmann machines. They also allowed for exciting collaborations with specialists in parallel computing facilities, such as Patrick Garda and Eric Bellhaire.
Today, the situation is radically different. Through cloud computing or multicore processors, researchers have access to sufficient computing power to do machine learning on large neural networks.
Why has France not translated these advances into industrial successes?
R.A.: Technology transfer has always been faster in the United States than in France. With some former DIAM staff members, I created two startups in 1999-2000: Miriad Technologies for online quality control of industrial processes and Sudimage for image analysis.
We soon had many prestigious customers (Airbus, Renault, Cnes, Snecma, CEA, Saint Gobain, Avantis, SAT, Air Liquide, etc.). I registered five patents in AI, which facilitated a first round of capitalization; but these startups didn’t have the financial means to compete with American growth companies. Google, for example, had several hundred full-time researchers after two years of existence... while I was mortgaging my house to launch Miriad Technologies and Sudimage. But that’s another story...
What are you doing now?
R.A.: I am a professor in Houston, where I joined my wife, an American pianist. My applied mathematics research focuses on 3D biomedical images, the emergence of rare events in genetic evolution, gene interactions in oncology, deep learning techniques, etc. The biomedical databases available in the US are growing at a rapid pace.
France is behind in generating this type of data and making it accessible. The key issue is still funding, and the ability to quickly set up scalable and high-capacity computing facilities, maintained by young computer scientists specialized in information technology.
Interview by Isabelle Bellin
Mathematics, Vision, Learning (MVA) Master’s programme
The MVA Master's has been France’s leading programme in the field for 20 years thanks to its grounding in top international research, its connection to an ecosystem of firms and startups—whose concerns provide material for the programme—and its ability to attract students from most of the grandes écoles in the country.
The objective of DAP (news media run through the ENS Paris-Saclay MVA) is to inform, clarify and foster debate around issues of data processing and analysis, from their foundations to their most promising scientific and economic prospects, thanks to the writing and rigour of our team of science journalists.
Topics covered include research, achievements, solutions, new uses, the people working in the field and their careers.
Our goal is to stimulate the ecosystem of academic research teams, companies—from startups to large international corporations—and teaching programmes (teaching staff, students and alumni).