BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//132.216.98.100//NONSGML kigkonsult.se iCalcreator 2.20.4//
BEGIN:VEVENT
UID:20260524T021851EDT-3557kJVKZG@132.216.98.100
DTSTAMP:20260524T061851Z
DESCRIPTION:Abstract\n\nLearning a categorical distribution from its sample
 s underlies many machine learning tasks\, including categorical generative
  modeling and classification.\n\nA critical property of a learned distribu
 tion is its calibration which measures how well the predicted probabilitie
 s reflect the true distribution. Proper calibration is essential both for 
 generative tasks\, where underconfidence or overconfidence can lead to und
 esirable effects\, and for classification with efficient inference\, where
  reliable confidence estimates are needed for cost-aware adaptive computat
 ion. However\, evaluating and learning calibrated models in the categorica
 l setting is challenging: the support space is typically prohibitively lar
 ge\, ground-truth probabilities are inaccessible\, and access to samples f
 rom the ground truth is usually very limited.\n\nThis thesis investigates 
 the problem of learning and evaluating calibrated models for categorical d
 ata\, addressing both generation and the setting of adaptive computation f
 or classification.\n\nFirst\, I propose a novel categorical generative mod
 el that combines a diffusion process with a structured\, sphere-packed enc
 oding and Gaussian mixture denoising\, which improves calibration\, sample
  quality\, and efficiency.\n\nSecond\, I develop a principled evaluation f
 ramework for categorical generative models based on synthetic distribution
  coarsening\, which provides interpretable diagnostics and statistical gua
 rantees in high-dimensional\, purely nominal categorical spaces. This cont
 ribution addresses the current gap in the literature on evaluation for cat
 egorical generative models\, as existing evaluation methods focus on conti
 nuous settings. Third\, I formulate a unified framework for adaptive compu
 tation with fixed classifiers\, showing that optimal resource allocation i
 n this setting reduces to accurately predicting per-sample error probabili
 ties. Lastly\, building on this\, I introduce a two-stage classification f
 ormulation in which multiple classifiers and the resource allocation modul
 e are trained jointly\, and provide a practical surrogate loss that is con
 sistent with a principled cost-aware target loss.\n\nTogether\, these cont
 ributions advance the understanding and design of calibrated probabilistic
  models and highlight the relationship between calibration\, evaluation\, 
 and adaptive computation in categorical settings.\n
DTSTART:20260114T150000Z
DTEND:20260114T170000Z
LOCATION:Room 603\, McConnell Engineering Building\, CA\, QC\, Montreal\, H
 3A 0E9\, 3480 rue University
SUMMARY:PhD defence of Florence Robert-Regol – Learning and evaluating cali
 brated models for categorical data
URL:/ece/channels/event/phd-defence-florence-robert-re
 gol-learning-and-evaluating-calibrated-models-categorical-data-370042
END:VEVENT
END:VCALENDAR