A multiperspective causal analysis of computing in predictive models based on machine learning
Résumé
Predictive models based on machine learning are more and more in use for different applications having social impacts: pattern (face) recognition, medical (cancer) diagnosis, predictive (crime) justice. For each, due to some kind of black-box effect, we are willing to ask the question: what caused the prediction? In this conference, I will show that it is fruitful - if not necessary - to adopt a multiperspective causal analysis for such predictive models. It is based on the assumption that for the same prediction device, different types of causes concurrently and convincingly can be advocated to have been at stake, but also that some specific causal profile - where some causes (material, formal, efficient, final) are found to be more present than others - can be attached to it. Of course, the whole causal enquiry cannot limit itself to the sole machine with its program but has to consider the whole system of human-human, human-idea, human-machine and machine-machine relationships. In this conference, I will not tackle the whole problem. I will focus this multiperspective causal analysis on some human-idea relationships. To show the fruitfulness of this approach, even at this restrictive level, it suffices to evoke the mathematical and metaphysical aspects of the debates around explainable artificial intelligence. As an example, I particularly will apply this causal profile analysis to the following competing mathematical approach in machine learning: 1) Bayesian, 2) Realist/Biomimetic (between brain and neural networks; see LeCun), 3) Causal statistics (Pearl, Schölkopf).
Origine :
Fichiers produits par l'(les) auteur(s)
Loading...