PhD defense of Marie GARIN

Marie GARIN

From federated learning towards a critical theory of heteromatic learning

Abstract

This manuscript attests to a significant epistemological swing. This document consists of two sections: the first relates to mathematics --- the initial discipline in which this thesis began --- and the second to Sciences and Technologies Studies (STS). The research object around which this manuscript revolves is machine learning --- described in the second section as heteromatic learning. The increasing digitization of societies and the consubstantial proliferation of digital traces --- whose data correlation constitutes the very heart of how heteromatic learning works --- have elevated surveillance and privacy issues to the rank of major societal concerns. Thus, the first section is devoted to an exploration of a machine learning method that does not centralize data and is therefore often presented as improving privacy~: federated learning. A hemistiche presents the genealogy of the epistemological swing and introduces the main methodological anchors guiding this approach; namely critical theory, some of Foucault's writings and STS. The second section, reflecting a reflexive and critical approach towards the first section, opens with an analysis of a controversy on the relationship between federated learning and privacy, and concludes with an examination of the political dimensions of the literature on technical means presented as solutions to privacy issues. This change of focus continues with a critical investigation of the dominant regimes of truth, addressing respectively the discrimination issues associated with heteromatic learning and the ecological issues of the digital device. Each of these two chapters concludes with the exploration of an alternative field of intelligibility; algorithmic equity in the light of neoliberal governmentality in the first case, and the study of the "environmental impacts" of the digital system in the light of Langdon Winner's critical philosophy of technology in the second.

Keywords

Machine learning, collaborative learning, critical theory, STS (Sciences and Technologies Studies), algorithmic fairness, privacy

Supervision

• Nicolas VAYATIS, PR, Centre Borelli, ENS Paris-Saclay (directeur de thèse)

Jury

• Francesca MUSIANI, Chargée de Recherche CNRS, CIS (rapportrice)
• Sébastien GAMBS, Professeur (rapporteur), UQAM
• Tiphaine VIARD , Maîtresse de conférence, Télécom Paris
• Anne-Laure LIGOZAT, Professoresse, LISN, ENSIIE
• Aurélien BELLET, Directeur de recherche, INRIA
• Pablo JENSEN, Directeur de recherche, ENS Lyon