LFI / TRAIL seminar “Generation of realistic and robust counterfactual explanations”

As part of the LFI LIP6 seminars and the joint AXA/Sorbonne University/LIP6 Trustworthy and Responsible AI Lab (TRAIL) laboratory, we will have the pleasure of welcoming Victor Guyomard (Orange Labs and University of Rennes) on Thursday June 22 at 2 p.m. for a talk entitled "Generation of realistic and robust counterfactual explanations".

Teaser:

The aim of the thesis is to explain individual decisions made by AI with a focus on counterfactual explanations. In this presentation, Victor will introduce two contributions:
1) the development of VCnet, a self-explanatory model that combines a predictor and a generator of counterfactuals that are learned simultaneously. The architecture is based on a variational autoencoder, conditioned on the output of the predictor to generate realistic counterfactuals (close to the distribution of the target class). VCnet is able to generate predictions as well as counterfactual explanations without having to solve another minimization problem.
2) the proposal of a new formalism, CROCO, to generate counterfactual explanations robust to the variation of the inputs of the counterfactual. This form of robustness involves finding a compromise between the robustness of the counterfactual and the proximity to the example to be explained. CROCO generates robust counterfactuals while effectively managing this trade-off and guaranteeing the user minimal robustness. Empirical evaluations on tabular datasets confirm the relevance and effectiveness of the proposed approach.

References: 

VCNet: A self-explaining model for realistic counterfactual generation. V. Guyomard, F. Fessant, T. Guyet, A. Termier, T. Bouadi.
Generating robust counterfactual explanations. V. Guyomard, F. Fessant, T. Guyet, T. Bouadi, A. Termier.

This seminar is organized jointly with the French Chapter of the IEEE Computational Intelligence Society