Wednesday 20th November 2019
Ground Floor Seminar Room
25 Howland Street, London, W1T 4JG
A walkthrough of advanced PAC-Bayes results
PAC-Bayes is a generic and flexible framework to address generalisation abilities of machine learning algorithms. It leverages the power of Bayesian inference and allows to derive new learning strategies. I will briefly present the key concepts of PAC-Bayes and focus on recent advances in high-dimensional ranking (https://arxiv.org/abs/1511.02729), non-iid learning (https://arxiv.org/abs/1610.07193), deep neural networks (https://arxiv.org/abs/1905.10259), and contrastive learning (https://arxiv.org/abs/1910.04464).
Benjamin Guedj is a Principal research scientist at University College London (Department of Computer Science), Inria and the Alan Turing Institute, leading the Inria@London initiative. He obtained his PhD in 2013 from Sorbonne Université, and his main lines of research are statistical learning theory, machine learning, deep learning and computational statistics. He is an expert in PAC-Bayes theory, which was the topic of his recent ICML 2019 tutorial with John Shawe-Taylor. Web: https://bguedj.github.io