Forum Numerica - The Bouncer and the Boundary: Modern Decision-Making Algorithms and Their Explanations

The impressive performance of neural-net based classifiers is now at the core of many decision-making systems. Despite their societal importance, these systems are still not well understood. Decisions are increasingly explained locally by their designers, often considering the decision-making system as a black-box. Decision boundaries are an interesting abstraction to reason about the output of these classifiers. Curiously, an increasing amount of private actors map this local explanation setup, but in destination to the user in order to convince him of the fair nature of the decision he faces. We will show how this could be seen as a fallacious move, as there is no way for an individual to verify the truthfulness of the provided explanations.

Erwan was a senior research scientist at Technicolor RsembioI (2009-2018), where he worked on scalable storage, processing and machine learning for data analytics and monitoring of home devices. He owns a PhD on distributed systems, and a background on graph mining algorithms. He obtained his habilitation (HDR) on November 2016 from University of Rennes 1. He is the president of the gozdata association, that provides the gozmail service, built from free software. In 2019, he joined the board of the Société Informatique de France, and in 2020 the scientific council of Inria's REGALIA. He is broadly interested by the impact of technology on society.

https://erwanlemerrer.github.io/