Working Group: Explainability and interpretability

Leader(s)

Brieuc Conan-Guez

Frédéric Pennerath

Mailing list

In order to get information about the WG activities, you can subscribe to its mailing list at https://sympa.inria.fr/sympa/info/mosaik-ia-explicable.

Description

The aim of the working group is to study and develop methods of explainability and interpretability for making eXplainable Artificial Intelligences (XAI). While these two terms of explainability and interpretability are generally used interchangeably, they refer to two slightly different meanings in the specific context of this working group: firstly, explainability is the ability of computational means to intelligibly describe AI decision-making to human beings. More generally, interpretability is the design of practical and theoretical tools that contribute to the understanding of the mechanisms underlying an AI, such as mechanistic interpretability, which aims to make explicit the latent concepts learned in LLMs.

The need for explainability is obvious and strategic, as it conditions the adoption of AI by society: the widespread use of AI depends entirely on the trust that human beings are willing to place in it and this trust can only come from control: to be entitled to make any decision with socio-economic impact, an AI must be able to justify at any time its decision with an explanation, in order to verify its consistency, relevance, ethical and legal character (absence of bias, etc.).

The interest in interpretability is of a different nature: by “deciphering” the internal logic of an AI, we can hope to improve it and ultimately advance research through a virtuous circle. In a way, interpretability can be broadly understood as the experimental science whose objects of study are artificial neural networks, analogous to neuroscience, which studies the brain using MRIs, or physics, which probe reality using new instruments. The tools available are both practical (i.e. software tools to aid analysis and visualization) and theoretical (recourse to statistics, information theory, statistical physics, etc.).

While the objectives of explainability and interpretability are different, in practice the two subjects are closely linked, insofar as an explainability technique can also often meet an interpretability objective, and vice versa.

Members

  • Bannay, Aurélie
  • Blansché, Alexandre
  • Bonnin, Geoffray
  • Boudjeloud-Assala, Lydia
  • Brun, Armelle
  • Castagnos, Sylvain
  • Cherif, Mohamed
  • Conan-Guez, Brieuc
  • D’Aquin, Mathieu
  • Dudyrev, Egor
  • Gendron-Audebert, Barbara
  • Hamadouche, Toufik
  • Kabrit, Yan
  • Labba, Chahrazed
  • Langlois, David
  • Lauzzana, Gabriel
  • Lieber, Jean
  • Parmentier, Yannick
  • Pennerath, Frédéric
  • Prouteau, Thibault
  • Zheng, Estelle
Back to top