Futoshi FURAMI – https://sites.google.com/view/futoshifutami/home (invité de Ikko Yamane)

Le vendredi 12 septembre 2025
de 11h à 12h (café d’accueil à partir de 10h40)
à l’ENSAI, Campus de Ker Lann à Bruz – salle 10

Title :Uncertainty Quantification through Calibration in Classification Problems

Abstract:In recent years, in high-risk domains such as medical diagnosis, weather forecasting, and financial risk management, there has been a strong demand for the reliability of probabilities output by machine learning models. A representative measure for assessing this reliability is calibration, which evaluates the degree of agreement between predicted probabilities and the actual frequency of events. However, it has been reported that modern deep learning models often exhibit poor calibration, and improving this issue has become an urgent challenge.
In this talk, I will introduce basic ideas related to calibration, as well as our recent research, including generalization error analysis of calibration errors based on statistical learning theory, and correction algorithms utilizing PAC-Bayesian methods.