BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Theoretical characterization of uncertainty in high-dimensional ma
 chine learning
DTSTART:20220831T090000
DTEND:20220831T110000
DTSTAMP:20260928T191308Z
UID:774e66ecfea829498b0f6c6972109749162666ad23956814c00fd0f5
CATEGORIES:Conferences - Seminars
DESCRIPTION:Lucas Clarte\nEDIC candidacy exam\nExam president: Prof. Nicol
 as Flammarion\nThesis advisor: Prof. Lenka Zdeborová\nCo-examiner: Prof. 
 Matthieu Wyart\n\nAbstract\nIn modern machine learning\, quantifying the u
 ncertainty\nof a model’s output is required to obtain reliable predictio
 ns\,\nespecially in sensitive applications like medical diagnosis.\nAmong 
 the three papers presented here\, one is concerned with\nstudying the asym
 ptotic behaviour of the uncertainty of logistic\nregression\, while the tw
 o others introduce approximate Bayesian\nmethods to improve uncertainty qu
 antification. The goal of our\nongoing and future research is to provide s
 tatistical guarantees\nfor the uncertainty of various algorithms\, in the 
 high-dimensional\nregime and in solvable models.\n\nBackground papers\n1) 
 Don't Just Blame Over-parametrization for Over-confidence: Theoretical Ana
 lysis of Calibration in Binary Classification (link: http://proceedings.ml
 r.press/v139/bai21c/bai21c-supp.pdf )\n2) A Simple Baseline for Bayesian U
 ncertainty in Deep Learning (link : https://proceedings.neurips.cc/paper/
 2019/hash/118921efba23fc329e6560b27861f0c2-Abstract.html )\n3) Being Baye
 sian\, Even Just a Bit\, Fixes Overconfidence in ReLU Networks (link : ht
 tps://proceedings.mlr.press/v119/kristiadi20a.html )
LOCATION:
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR