Operational AI Ethics

News

L’explicabilité des algorithmes au service de la lutte contre le blanchiment d’argent

Digital Trust, Data Science & AI — 27/10/20

D’après les estimations d’Europol, 200 milliards d’euros de fonds criminels circulent chaque année en Europe[1], et seulement 1% de ces flux financiers illicites sont saisis. Les systèmes de lutte anti-blanchiment actuels ont donc une grande marge de progression, notamment pour pouvoir contrer des méthodes criminelles de plus en plus sophistiquées. L’IA s’avère d’une grande aide à l’amélioration du système de lutte anti-blanchiment mais, en parallèle, elle engendre de nouvelles formes de risques pour les individus. C’est là que l’explicabilité intervient.

Lire la suite…Read more…

Do machines understand what they read?

Digital Trust, Data Science & AI — 22/10/20

One of the challenges of natural language processing (NLP) is to get the meaning right. Subtle changes in a sentence can radically change its meaning. Computers may have difficulty detecting irony and sarcasm, distinguishing an affirmation of fact from an opinion. In their recent paper, Fabian Suchanek from Télécom Paris and his coauthor Mechket Emna Mahouachi from ENSTA provide an introduction to the fascinating field of information extraction. Professor Suchanek is directing a research program at Télécom Paris on the subject of extracting complex meaning from text.

Lire la suite…Read more…

Explainability and fundamental rights

Digital Trust, Data Science & AI, Faculty Members — 16/10/20

When does the law require an algorithmic decision to be explained? A number of laws (GDPR, competition law, consumer protection) require companies to be transparent, fair and provide explanations for decisions they make. These obligations are usually justified because of asymmetries in market power or information. For public administrations, the obligation to provide explanations has constitutional roots.

Lire la suite…Read more…

Pourquoi les algorithmes de reconnaissance faciale ne peuvent pas être parfaitement équitables

Digital Trust, Data Science & AI, Faculty Members — 06/10/20

En juin 2020, un algorithme de reconnaissance faciale a conduit à l’arrestation indue de Robert Williams, un Africain-Américain, pour un crime qu’il n’avait pas commis. Suite à un vol à l’étalage dans un quartier huppé de Detroit, dans le Michigan, le malfaiteur apparaissant dans une vidéo floue a en effet été incorrectement identifié comme Robert Williams, avec la photo de son permis de conduire. Après plusieurs heures, la police l’a relâché et a présenté ses excuses, mais cet épisode soulève de sérieuses questions quant à la pertinence des algorithmes de reconnaissance visuelle. La partie la plus troublante dans cette histoire est qu’il a été montré que les algorithmes de reconnaissance faciale sont moins précis pour les visages noirs que pour les blancs. Mais pourquoi est-ce le cas et que peut-on faire contre cela ?

Lire la suite…Read more…

To simplify AI regulation, use the GDPR’s high-risk criteria

Digital Trust, Data Science & AI, Faculty Members — 22/09/20

There are a multitude of contexts where we would expect an AI application to be regulated to protect users. The European Commission proposes a single AI regulation and a binary test for applying it: if the AI application is high-risk, then specific AI regulations apply. If not, then no specific AI regulations apply. The Commission proposes a definition of AI and two cumulative criteria for determining if an application is high risk. This two-criteria test gives the appearance of legal certainty, but it’s likely to create the opposite, for two reasons.

Lire la suite…Read more…

The Operational AI Ethics program allows us to deal with the subject of bias and the explicability of algorithms in all its diversity, by bringing together complementary expertise well beyond our respective disciplines.

Stephan Clémençon, professor in applied mathematics and machine learning at Télécom Paris

Visit our blog

An interdisciplinary team

Our interdisciplinary team federates six academic fields (applied math, statistics, computer science, economics, law and sociology) around the ethical issues raised by artificial intelligence.

What our researchers say

To be reliable, a decision system must meet criteria of explainability in addition to performance criteria. We seek to integrate these criteria when learning these systems.

Florence d'Alché-Buc, professor in massive data processing and mining at Télécom Paris and holder of the Data Science and Artificial Intelligence for Digitalized Industry and Services chair

Hybrid approaches make it possible to combine symbolic AI (based on knowledge representation) and machine learning (especially from data) leading to a better explicability of algorithmic decisions.

Isabelle Bloch, professor in applied mathematics and artificial intelligence, in particular for spatial reasoning, at Télécom Paris

The European Commission and the regulators all want explainable algorithms, but the devil is in the detail. The questions that need to be answered are: for whom and why?

Winston Maxwell, director, law and technology studies at Télécom Paris

If, the explicability of algorithms is not necessarily necessary for an algorithm which plays chess, it becomes essential when it is about a critical system.

David Bounie, head of the Department of Economics and Social Sciences at Télécom Paris

PhD Students

Five research pillars

Teaching and research chairs

Training offers

The production of data from networks, connected objects, sensors or industrial processes has largely contributed to the digital transition of companies, which now consider data as a resource in its own right. All of a company’s activities and jobs have now been impacted, resulting in a growing need for training for teams and managers. Télécom Paris offers a variety of training options tailored to meet all needs.

Post-Master’s Degrees

Specialized Studies Certificates

Télécom Évolution offers two specialized studies certificates, Data Scientist and Artificial Intelligence, both of which are delivered by Télécom Paris, with contributions from ENSTA Paris for the second certificate. In collaboration with École polytechnique d’Assurance, it offers an Executive MBA Data Scientist for Insurance Careers.

Télécom Évolution also offers many inter-company training courses as well as the MOOC, The Basics of Big Data, which enrolls 7,000 learners every session.

Short training

Télécom Evolution offers the short training course The ethical challenges of artificial intelligence. This 2-day training course provides an overview of the main ethical, legal and societal issues surrounding AI.

Learn more about Télécom Évolution