BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:ECAS-ENBIS Course: Conformal Prediction: How to quantify uncertain
ty of machine learning models?
DTSTART:20230910T120000Z
DTEND:20230910T160000Z
DTSTAMP:20241007T015800Z
UID:indico-event-41@conferences.enbis.org
DESCRIPTION:\n\nECAS-ENBIS Course: Conformal Prediction: How to quantify u
ncertainty of machine learning models?\n\nPart of the ENBIS-23 Valencia
conference.\n\nThis half-day course is a joint initiative from ENBIS and E
CAS (http://ecas.fenstats.eu/) which has provided courses since 1987 to ac
hieve training in special areas of statistics for both researchers and tea
chers for universities and professionals in industry fields.\n\nInstructor
\n\nMargaux Zaffran (INRIA\, Ecole Polytechnique and EDF\, France)\n\nOver
view\n\nBy leveraging increasingly large data sets\, statistical algorithm
s and machine learning methods can be used to support high-stakes decision
-making problems such as autonomous driving\, energy\, medical or civic ap
plications\, and more. In order to ensure the safe deployment of predictiv
e models\, it is crucial to quantify the uncertainty of the resulting pred
ictions\, communicating the limits of predictive performance. Therefore\,
uncertainty quantification attracts a lot of attention in recent years\, p
articularly methods that are based on Conformal Prediction. Conformal Pred
iction provides controlled predictive regions for any underlying predictiv
e algorithm (e.g.\, neural networks and random forests)\, in finite sample
s with no assumption on the data distribution except for the exchangeabili
ty of the train and test data. Conformal Prediction has already been succe
ssfully used to predict in real time the results of the last US presidenti
al elections (2020) by the Washington Post.\n\nThis short course is a firs
t introduction to Conformal Prediction\, aimed at a broad audience of prac
titioners and researchers. It requires basic knowledge in statistics\, pro
bability and machine learning (including regression\, classification\, tra
ining and validation splitting strategies\, for example\, but no prior kno
wledge in any specific machine learning algorithms is required). Availab
le software and implementations of Conformal Prediction will be reviewed.\
n\nWe will introduce the Conformal Prediction framework\, in its generic v
ersion\, applicable to both regression and classification tasks. Then\, we
will review the existing challenges\, such as the computational cost\, co
nditional and adaptive coverage or even distribution shifts. and the recen
t solutions that have been proposed to handle them. Finally\, we will focu
s on one specific challenge: time series forecasting. Temporal data are no
t exchangeable\, therefore they do not met the only assumption required by
Conformal Prediction. We will highlight new developments on this topic\,
and illustrate these procedures on the task of forecasting French electric
ity prices.\n\nOutline\n\nBelow is given an outline of the course\, with t
he main references for each part.\n\nPart 1: Introduction to conformal pre
diction\n\n\n Reminders on machine learning\, in particular supervised lea
rning\, and dive into quantile regression\n Split conformal prediction: fr
om standard regression to classification\n A generalized framework: full c
onformal prediction\n\n\n\nPart 2: Challenges\n\n\n Tradeoff between compu
tational cost and statistical efficiency: jackknife+ and CV+ approaches\n
Improving the adaptivity of conformal prediction\n Beyond the exchangeabil
ity assumption\n\n\nPart 3: Adaptive Conformal Predictions for Time Serie
s\n\nShort bio\n\nMargaux is a last year PhD student at both INRIA and Eco
le Polytechnique (Paris)\, with Aymeric Dieuleveut (Ecole Polytechnique)
and Julie Josse (INRIA). She also works with Electricité de France\, t
he leading producer and supplier of electricity in France and Europe\, in
collaboration with Olivier Féron and Yannig Goude.\n\nHer research int
erests lie in developing tools able to quantify uncertainty with any machi
ne learning algorithm\, and more generally\, in probabilistic predictions\
, motivated by real-world applications in the energy sector. This specifit
y attracts her towards time series and missing data. Before focusing on co
nformal prediction and distribution-free uncertainty quantification\, she
analyzed quantile regression approaches\, based on generalized additive mo
dels\, at the University of Bristol.\n\nhttps://mzaffran.github.io\n\n\n\n
https://conferences.enbis.org/event/41/
URL:https://conferences.enbis.org/event/41/
END:VEVENT
END:VCALENDAR