CS Colloquium: Matteo Sesia (USC Marshall School of Business) - Conformal inference for uncertainty-aware classification

Speaker: Matteo Sesia, USC Marshall School of Business

Talk Title: Conformal inference for uncertainty-aware classification

Series: Computer Science Colloquium

Abstract: Complex machine learning classifiers, including deep neural networks, are sometimes able to achieve very high predictive accuracy, but they are not designed to realistically capture uncertainty or to estimate reliable probabilities. In fact, these models are often overconfident, and this issue can make it challenging for practitioners to accept the use of machine learning algorithms in delicate real-world applications. This talk will describe recent advances in the field of conformal inference which allow us to address the overconfidence of machine learning classifiers. First, this talk will present a powerful and statistically principled methodology for assessing the uncertainty of predictions computed by any pre-trained classification model, in such a way as to account for possible heterogeneity in the levels of uncertainty affecting different individual data points. Then, building upon the previous results, this talk will present a novel methodology for training deep neural networks in such a way as to learn multi-class classification models that are less prone to overconfidence, ultimately leading to even more reliable uncertainty-aware predictions.

Prof. Sesia will give his talk in person at RTH 115 and we will also host the talk over Zoom.

Join Zoom Meeting
https://usc.zoom.us/j/92821217575
Meeting ID: 928 2121 7575

This lecture satisfies requirements for CSCI 591: Research Colloquium.

Biography: Matteo Sesia is an assistant professor in the department of Data Sciences and Operation at the USC Marshall School of Business.
Matteo joined USC Marshall in 2020, immediately after earning a PhD in Statistics from Stanford University, where he was advised by Emmanuel Candes. Matteo's research primarily focuses on developing novel methodology for model-free statistical inference with big data, and on developing statistically principled algorithms for uncertainty-aware machine learning.

Host: Yan Liu

Webcast: https://usc.zoom.us/j/92821217575

Location: Ronald Tutor Hall of Engineering (RTH) - 115

WebCast Link: https://usc.zoom.us/j/92821217575

Audiences: Everyone Is Invited

Contact: Department of Computer Science