BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//CERN//INDICO//EN
BEGIN:VEVENT
SUMMARY:Class conditional conformal prediction for multiple inputs by p-va
 lue aggregation
DTSTART:20251106T100000Z
DTEND:20251106T111500Z
DTSTAMP:20260504T170900Z
UID:indico-event-14999@indico.math.cnrs.fr
DESCRIPTION:Speakers: Jean Baptiste Fermanian (INRIA)\n\nConformal predict
 ion methods are statistical tools designed to quantify uncertainty and ge
 nerate predictive sets with guaranteed coverage probabilities. This work 
 introduces an innovative refinement to these methods for classification ta
 sks\, specifically tailored for scenarios where multiple observations (mu
 lti-inputs) of a single instance are available at prediction time. Our ap
 proach is particularly motivated by applications in citizen science\, whe
 re multiple images of the same plant or animal are captured by individual
 s. Our method integrates the information fromeach observation into conform
 al prediction\, enabling a reduction in the size of the predicted label s
 et while preserving the required class-conditional coverage guarantee. Th
 e approach is based on the aggregation of conformal p-values computed fro
 m each observation of a multi-input. By exploiting the exact distribution
  of these p-values\, we propose a general aggregation framework using an 
 abstract scoring function\, encompassing many classical statistical tools
 . Knowledge of this distribution also enables refined versions of standar
 d strategies\, such as majority voting. We evaluate our method on simulat
 ed and real data\, with a particular focuson Pl@ntNet\, a prominent citize
 n science platform that facilitates the collection and identification of p
 lant species through user-submitted images. Paper\n\nhttps://indico.math.c
 nrs.fr/event/14999/
LOCATION:Auditorium 3 (Toulouse School of Economics)
URL:https://indico.math.cnrs.fr/event/14999/
END:VEVENT
END:VCALENDAR