PDIA’21

« I will present part of the work carried out in the CRIL lab, within the framework of the ANR « EXPEKCTATION » chair, which got started 6 months ago. Our purpose is to leverage the concepts and algorithms developed in the knowledge representation (KR) field to address eXplainable AI (XAI) issues. The focus will be on classifiers, and more precisely on defining XAI queries for classifiers, identifying their computational complexity (and suggesting how to deal with their intractability), as well as on the computation of intelligible explanations from the predictions made. »

Les explications fournies dans le cadre de l’apprentissage automatique peuvent se porter à plusieurs niveaux. Cet exposé s’intéresse d’abord à la compréhension des données elles-mêmes, dans un but d’analyse exploratoire et de description intelligible, qui s’inscrit dans les problématiques de type data-to-text : l’objectif est de permettre à un utilisateur de comprendre le contenu des données en les résumant par le biais de formulations linguistiques, ce qui soulève en particulier les problèmes de choix des mots et de cohérence des résumés. Le second niveau considéré est celui de tâches de classification, dans le cadre classique de l’interprétation locale, post-hoc et agnostique de la prédiction d’une classe pour une donnée. Les questions soulevées sont celles des risques liés à la définition de la localité et à la génération d’explications non justifiées.

« My presentation will cover AI Explainability from a law and social sciences perspective, including the following points:
– What purposes do explanations serve?
– What audience(s)?
– What costs and benefits, and when are explanations worth the costs?
– What legal texts impose algorithmic explanations?
– How can explanations help mitigate human biases (such as automation bias)? »

Algorithmic decisions are now being used on a daily basis, and based on Machine Learning (ML) processes that may be complex and biased. This raises several concerns given the critical impact that biased decisions may have on individuals or on society as a whole. Not only unfair outcomes affect human rights, they also undermine public trust in ML and AI. In this talk, we will address fairness issues of ML models based on decision outcomes, and we will show how the simple idea of feature dropout followed by an ensemble approach can improve model fairness without compromising its accuracy. To illustrate we will present a general workflow that relies on explainers to tackle process fairness, which essentially measures a model’s reliance on sensitive or discriminatory features. We will present different applications and empirical settings that show improvements not only with respect to process fairness but also other fairness metrics.

The adoption of machine learning in critical contexts requires a reliable explanation of why the algorithm makes certain predictions. To address this issue, many methods have been proposed to explain the predictions of these black box models.
An overview of the explicability methods will be presented first. Following this presentation, a focus will be done on the measure of “performance” of these explanations. While it is possible to establish a number of desirable properties of a good explanation, it is more difficult to evaluate them. As a result, no measures are actually associated with the properties of the consistency and the generalization of explanations. We will introduce a new procedure to compute these two new measures, Relative Consistency ReCo and Mean Generalization MeGe, respectively for consistency and generalization of explanations. We will demonstrate the potential of the measures by applying them to different families of models, revealing an interesting link between gradient-based explanations methods and 1-Lipschitz networks.