BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Should Ensemble Members Be Calibrated? - Dr Xixin Wu\, Cambridge U
 niversity Engineering Department
DTSTART:20210216T120000Z
DTEND:20210216T130000Z
UID:TALK156793@talks.cam.ac.uk
CONTACT:Dr Kate Knill
DESCRIPTION:*Abstract:* Underlying the use of statistical approaches for a
  wide range of applications is the assumption that the probabilities obtai
 ned from a statistical model are representative of the “true” probabil
 ity that event\, or outcome\, will occur.  Unfortunately\, for modern deep
  neural networks this is not the case\, they are often observed to be poor
 ly calibrated. Additionally\, these deep learning approaches make use of l
 arge numbers of model parameters\, motivating the use of Bayesian\, or ens
 emble approximation\, approaches to handle issues with parameter estimatio
 n. This paper explores the application of calibration schemes to deep ense
 mbles from a theoretical perspective. The theoretical requirements for cal
 ibration\, and associated calibration criteria\, are first described. It i
 s shown that well calibrated ensemble members do not necessarily yield a w
 ell calibrated ensemble prediction. Furthermore if the ensemble prediction
  is well calibrated then its performance cannot exceed that of the average
  performance of the calibrated ensemble members.  Empirical results on CIF
 AR-100 are used to support these theoretical developments. Additionally th
 e relationships between ensemble calibration for classification and regres
 sion are discussed.\n\n*Bio:*\nXixin Wu is a Research Associate in the Spe
 ech Group of the Machine Intelligence Laboratory\, Engineering Department 
 of Cambridge University.  He obtained his PhD degree from The Chinese Univ
 ersity of Hong Kong.  His research interests include speech recognition an
 d synthesis\, speaker verification and neural network uncertainty. Xixin i
 s a member of IEEE and ISCA.
LOCATION:Zoom: https://zoom.us/j/95352633552?pwd=RzJVK2UzOGZyNU5mVHd1Y1VPT
 2tDUT09
END:VEVENT
END:VCALENDAR