Martin Vechev, Professor of Computer Science at ETH Zurich
Title: Beyond Accuracy: Towards Robust, Fair and Certified Deep Learning
Host: Zhong Shao
The desire to deploy deep learning models in the wild has led to a fundamental challenge: the need to train and certify models that not only enjoy high accuracy on a dataset but also (provably) satisfy critical domain-specific properties (e.g., fairness, robustness, safety, etc.). Perhaps surprisingly, creating such models has turned out to be substantially harder than expected, pointing to key limitations of existing learning techniques and dictating the need for methodological advances beyond more labeled data or compute power.
In this talk I will discuss some of the recent progress we have made on creating deep learning models which come with provable mathematical guarantees. This progress is largely based on new methods sitting at the intersection of different areas including symbolic reasoning, statistics, machine learning, probabilistic programming, and others. In the process I will also elaborate on several open challenges and exciting research opportunities which if solved, will enable broader and safer use of machine learning.
Martin Vechev is a Professor of Computer Science at ETH Zurich. Prior to ETH, he was a Researcher at the IBM T.J. Watson Research Center in New York, USA. He received his PhD from Cambridge University, UK. His main research interests sit at the intersection of machine learning, probabilistic reasoning and programming languages. His work has received various recognitions including the ACM SIGPLAN Robin Milner Young Researcher Award, ERC Starting Grant, SIGPLAN and CACM Research Highlights, and many others. Beyond academia, he has co-founded 3 startups: two acquired in 2020, with the latest LatticeFlow focusing on robust and trustworthy AI. He has also spoken at the European Parliament on topics related to AI and democratizing scientific innovation.