| Véronique Fuller | Avril 2019 |
Un récent rapport  suggérait que si en Europe une start-up sur 12 est référencée en intelligence artificielle (IA), seules 60% d’entre elles apporteraient vraiment la preuve qu’elles utilisent l’IA. Que recouvre ce domaine à la croisée de plusieurs disciplines ?
Michalis Vazirgiannis, Professeur et Chercheur au Laboratoire d’Informatique de l’École polytechnique (LIX) nous parle de la représentation des données en graphe, des applications de l’IA de la reconnaissance d’emails spams à la génération automatique d’un compte-rendu de réunions, de l’algorithme de Google qui traite 4 millions de demandes par minute et aussi de la vie sur le campus, qui accueille désormais 40% d’étudiants internationaux, et où une part croissante des cours est enseignée en anglais.
How would you define “Artificial Intelligence”, “Machine learning”, and “deep learning”?
Artificial intelligence is a broad area involving methods and techniques that allow computers to extract patterns from data to build new knowledge potentially extending previous ones, also making deductions, inferences and predictions. AI started decades ago, but due to the lack of enough Data, industrial interest and computing power, it didn’t truly take up until a few years ago.
Machine learning, part of AI, involves the methods and processes where the computer learns from some previous experience (for instance a bunch of emails characterized as spam) and learns a model that is able to predict classification for unknown data, for instance to decide if a new email should be classified as spam or not. To put it more formally, it is the process of fitting a mathematical function to a training set and then in most cases we use this function to predict the classification of new unclassified data. The implications of machine learning in current systems and applications are numerous and affect almost every aspect of socioeconomic and personal life.
The main changing factor on the impact of AI in everyday life is the development and dominance of deep learning over the last few years. Deep learning is a set of methods and systems that capitalize on potentially complex architectures involving neural networks. Neural networks are layers of units – called “perceptrons” – that given a certain set of inputs, output a binary decision based on nonlinear functions. Given a set of inputs and a desired output, neural networks learn the potentially complex function that maps input to output. A prominent application is a summarization, for example, when given a document as input, we get a summary of fixed length as output.
When using a search engine like Google, we are all using ranking algorithms. How has machine learning made it evolve in recent years?
Web search is one of the dominant applications of digital technology – there are about 4M queries per minute submitted to Google – the most successful search engine to date. The challenge is how to select the top 10 answers to the query – among the million matching ones – to display on the first page of results (it is well known that most people don’t go beyond the first page of results). Traditionally, only text matching was used, but this left internet users quite vulnerable to spamming. In the 90’s, the founders of Google invented a graph-based algorithm that measured the importance of a web page in the web graph based on its incoming links. This algorithm – known as “pagerank” – contributed to better ranking as it was jointly taken into account on top of the text based similarity among query and pages to produce the rankings. This gave Google results their well-known quality and practically dominated the web search and advertising market in the Western world. Of course, now ranking is a more complex process where hundreds of criteria plus human supervision are taken into account.
More recent machine learning algorithms contributed to better rankings, as the ranking is learned from training data. Moreover, our behavior, which is heavily tracked, combined with other data, is used to produce user models that are exploited in web marketing and advertising for targeted recommendations, influence maximization and so on.
In addition, AI is driving new application domains such as medical diagnosis, autonomous vehicles, risk management, legal informatics, and more.
Many École Polytechnique Professors teach in both the Cycle Ingénieur and the Bachelor Program. How is the presence of these young students changing the atmosphere of the campus?
A major part of the students are selected from schools all over the world – bringing thus a diverse and multicultural atmosphere to École Polytechnique. Between its various undergraduate and graduate-level programs (Bachelor of Science, Ingénieur Polytechnicien Program (Master’s level program), Masters of Science & Technology, and PhD), the school offers a variety of international programs and attracts a growing number of foreign students as well as researchers from around the globe (currently 40% of students and 39% of faculty members). For a closer look within the programs, the Bachelor of Science program is made up of 70% international students, the Masters of Science & Technology feature 60% international students and in the Cycle Ingénieur Polytechnicien, 25% of students are international students. This definitely contributes to the formation of a genuinely international environment where ideas and knowledge culminate upon the traditional excellence of l’X. Moreover, the English language used for teaching and everyday interactions at the school help to amplify this potential.
What advice would you give today to a secondary school student?
The world is quite more complex today than ever before. Students that want to be well placed in the societies of tomorrow need to be equipped with the capacity to adapt and change. I think that it is essential to invest in computer science skills with a sound quantitative background, but also soft skills like presentations, team building and participation. It is also very important to learn how to integrate and contribute in teams – this helps in developing leadership skills as well. Language skills are also essential as they offer access to the world’s populations, cultures and economies. Finally, learning and understanding history and philosophy enables global understanding of the current state of the world.
And more specifically to someone who is interested in Artificial Intelligence?
It is essential that, from an early age, students interested in this area start to get familiar with concepts such as algorithms, computational complexity, data structures and more. AI at scale involves building software systems as well – so software development skills (starting with python) are quite essential. It is also useful to build a sound foundation on applied math (i.e. probability, statistics, algebra, and optimization). I think that already understanding the basic concepts of machine learning would be beneficial. A lot of online courses of high quality already exist in order to start preparing students at a young age.