Seminar Series: Uncertainty Estimation for Deep Neural Networks
March 23 @ 2:00 pm – 3:00 pm
Danilo Silva is an Associate Professor in the Department of Electrical and Electronic Engineering at the Federal University of Santa Catarina, Brazil, where he leads the Machine Learning and Applications Research Group. He has also been a Visiting Researcher at the TAMIDS Scientific Machine Learning Lab since January 2026. He received his Ph.D. degree in Electrical Engineering from the University of Toronto in 2009 and held postdoctoral positions at the University of Toronto, at the École Polytechnique Fédérale de Lausanne, and at the State University of Campinas.
Location: Blocker 220 and Zoom
Zoom ID: 974 9688 4861
Passcode: 923446
Uncertainty Estimation and Selective Classification for Deep Neural Networks
Despite the impressive predictive performance of deep neural networks across diverse tasks, their predictions remain prone to errors, posing significant challenges in safety-critical applications. This talk discusses uncertainty estimation and its use in selective prediction, where a model abstains from low-confidence predictions to improve performance. In the first part, we examine selective prediction in standard multi-class classification, focusing on confidence estimation methods and post-hoc techniques that improve the risk-coverage trade-off.
We further reveal how prevalent regularization techniques, such as label smoothing, can improve classification accuracy while degrading selective classification by distorting probabilistic confidence estimates. In the second part, we address selective prediction for semantic segmentation, with emphasis on medical imaging. We derive an ideal image-level confidence estimator based on the Dice metric and introduce Soft Dice Confidence (SDC), a practical approximation with tight theoretical guarantees. Experiments on synthetic and medical datasets show that SDC yields superior confidence estimates even under distribution shift, substantially improving the risk-coverage trade-off for image-level abstention.