BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Date iCal//NONSGML kigkonsult.se iCalcreator 2.20.2//
METHOD:PUBLISH
X-WR-CALNAME;VALUE=TEXT:TSE
BEGIN:VTIMEZONE
TZID:Europe/Paris
BEGIN:STANDARD
DTSTART:20241027T030000
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
RDATE:20251026T030000
TZNAME:CET
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20250330T020000
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
UID:calendar.136962.field_date.0@www.tse-fr.eu
DTSTAMP:20260722T142456Z
CREATED:20250305T161001Z
DESCRIPTION:Bryan Wilder (Carnegie Mellon University)\, “Machine learning t
 o complement human decision making”\, Digital Workshop\, TSE & IAST\, Apri
 l 16\, 2025\, 12:30–13:30\, Auditorium A4.\n\nIn many settings\, machine l
 earning models are used in conjunction with a human decision maker. For ex
 ample\, consider a model used to make diagnoses in healthcare. Far from ha
 ving the model operate autonomously\, human clinicians may provide a secon
 d opinion on hard cases\, use the predictions to decide on followup action
 s\, and so on. This raises a challenge: how can we design machine learning
  models which best complement the strengths\, weaknesses\, and decision-ma
 king process of humans? I will start by discussing the triage setting\, wh
 ere the model must decide whether to query a (costly) human expert for hel
 p with a given example. We find that explicitly modeling the human's respo
 nse\, and adapting the model to reflect their strengths and weaknesses\, y
 ields better team performance than a model trained for accuracy in isolati
 on. Then\, I will turn to uncertainty quantification\, where the model out
 puts a set of possible labels for use by a human instead of just a point p
 rediction. Standard methods for producing prediction sets focus on ensurin
 g coverage of the true label\, agnostic to how useful a set is for a downs
 tream decision maker. We propose a method to optimize prediction sets rela
 tive to a utility function that models a human's decision-making process\,
  for example preferring sets that imply similar followup actions for clini
 cian. Empirically\, we find that our method produces prediction sets that 
 are more cohesive and clinically interpretable while retaining the accurac
 y guarantees as before.
DTSTART;TZID=Europe/Paris:20250416T133000
DTEND;TZID=Europe/Paris:20250416T143000
LAST-MODIFIED:20260119T111001Z
LOCATION:TSE & IAST\, April 16\, 2025\, 12:30–13:30\, Auditorium A4
SUMMARY:Digital Workshop
URL;TYPE=URI:https://www.tse-fr.eu/seminars/2025-machine-learning-complemen
 t-human-decision-making
END:VEVENT
END:VCALENDAR