Programme

Jeudi 16 Janvier 2025, 9h00-17h00, Amphi Hermite, Institut Henri Poincaré

09h00-09h30 Accueil

09h30-11h00 Mini-cours de Christophe Giraud : Introduction to fairness in statistical learning [slides]

Fairness of decisions based on output of machine learning algorithms is of paramount importance, in particular in some sensitive cases such as job hiring, decision in justice, admission in college, etc. Applying blindly standard ML technics can lead to unfair and discriminatory decisions. This mini-course will present the problematic and the origins of unfairness in statistical decision making. Different statistical approaches and technics for handling fairness issues will be introduced and discussed.

11h00-11h30 Pause Café

11h30-12h30 Exposé de Christophe Denis : Fairness guarantees in multi-class classification [slides]

Algorithmic Fairness is an established area of machine learning, willing to reduce the influence of hidden bias in the data. Yet, despite its wide range of applications, very few works consider the multi-class classification setting from the fairness perspective. In this talk, we focus on this question and extend the definition of approximate fairness in the case of Demographic Parity to multi-class classification. We specify the corresponding expressions of the optimal fair classifiers and describe a post-processing approach based on the plug-in principle to estimate the optimal predictor. We also illustrate the performance on both simulated and real data.

12h30-14h Pause Déjeuner

14h-15h30 Mini-cours de Jean-Michel Loubes : A mathematical point of view of  legal compliance: from audit to bias mitigation

We provide an in-depth analysis of the implications of the European AI regulation on mathematicians involved in designing AI algorithms. Specifically, we outline the key requirements for achieving legal compliance in machine learning systems and detail the necessary controls these algorithms must undergo. Furthermore, we explore mathematical
approaches to enforce fairness constraints at both group and individual levels, with examples using optimal transport theory. Finally, we examine methodologies for designing effective audit frameworks to assess algorithmic conformity with regulatory standards.

15h30-16h30 Exposé de Fanny Jourdan : L’explicabilité dans le NLP : identifier les biais dans les gros modèles de langue [slides]

Les grands modèles de langue, comme BERT pour la classification ou GPT pour la génération, ont révolutionné le traitement automatique des langues (NLP). Toutefois, leur complexité et leur opacité rendent difficile la détection et l'analyse des biais qu'ils peuvent reproduire ou amplifier. Contrairement aux approches classiques développées pour des tâches simples sur données tabulaires, de nouvelles méthodes d'explicabilité (XAI) adaptées aux modèles de NLP sont nécessaires pour comprendre ces phénomènes. Cette présentation propose un tour d'horizon des techniques modernes d'explicabilité permettant d'identifier, analyser et mieux appréhender les biais dans les représentations et les prédictions des grands modèles de langue.