Bio

I am an associate professor at the Korteweg-de Vries Institute for Mathematics at the University of Amsterdam, where I work on the mathematical foundations of machine learning. My current focus is to build up mathematically rigorous theory for explainable AI, the field that aims to explain the decisions of black-box machine learning systems to a user. My previous main research theme was adaptive methods for online learning and optimization, but I have also worked on topics in information theory, model selection for Bayesian statistics, and PAC-Bayesian concentration inequalities. My current work is funded by a VICI grant from the Dutch Research Council, which previously also awarded me VIDI, TOP2 and Rubicon grants. I am on the board of directors for the AI & Mathematics (AIM) initiative and chair the research committee of Platform Wiskunde Nederland (PWN). From 2023 to 2026 I was a member of the board of directors for the Conference on Learning Theory (COLT).

Initials and title for official stuff: Dr. T.A.L. van Erven.

Research

Machine learning systems tune millions or even billions of internal parameters to optimize their performance on a task, but our ability to explain how they arrive at their decisions has lagged far behind. The field of explainable machine learning aims to fill this gap, yet after more than a decade of intensive research it often remains unclear what exactly explanation methods are for and when they can be relied upon. My aim is therefore to build up formal mathematical foundations for explainable machine learning, moving from intuitive motivations and proof-of-concept demonstrations to precise definitions and strict, falsifiable claims. I discern three high-level themes: explanations must be succinct to be understandable, so how do we restrict them to be succinct? Formal definitions are key, so how do we choose them? And we can only trust explanations if we trust the method that produced them, so which guarantees can we give about explainability methods?

My previous main research theme was the design of mathematically well-founded machine learning methods for online convex optimization that work well out of the box, without any manual fine-tuning. The approach is to identify statistically ‘easy’ situations (e.g. low noise, small norm optimal parameters, etc.) in which it is possible to learn more efficiently, with less data, and then to construct adaptive methods that exploit these easy cases when present, but automatically fall back to slower robust learning strategies when there is no easy structure. This is technically challenging, because it typically requires automatic hyper-parameter tuning, for which we developed new approaches.

In the past I have also worked on topics in information theory, model selection for Bayesian statistics, and PAC-Bayesian concentration inequalities.

See also my publications and my blog.

Background

Previous Positions

I studied Artificial Intelligence at the University of Amsterdam and obtained my PhD in 2010 at the CWI in Amsterdam under the supervision of Peter Grünwald. I then held postdoc positions with Peter Grünwald at CWI, with Aad van der Vaart at the VU, and, on a Rubicon fellowship, with Pascal Massart at the Université Paris-Sud, where I was also a member of the INRIA team SELECT. Before moving to Amsterdam in 2020, I was an assistant professor in the statistics group at Leiden University.

Reviewing

I have been a senior program committee member for COLT since 2015 and was its open problems chair in 2023 and 2024. I was a senior program committee member for ALT in 2021 and 2023, and an area chair for NeurIPS from 2018 to 2020. I regularly review for COLT, NeurIPS, ICML, IEEE Transactions on Information Theory, the Journal of Machine Learning Research, Machine Learning, BNAIC/BeneLearn and others, and I regularly serve on PhD committees. I am a member of the Amsterdam ELLIS Unit.