Rabanus Derr

A smiling Rabanus in its natural habitat.

Hi there, I'm Rabanus, a PostDoc in the group "Foundations of Machine Learning Systems" led by Bob Williamson, and penguin-enthusiast (when you want to see some come to one of my talks ;) ). My research interest is best summarized as "understanding the statistical nature of machine learning".

We use machine learning tools in a variety of ways, from language generation to predicting job performance. People, individuals, groups, we, you and I are affected by those uses, in positive (e.g., useful) and negative (e.g., unfair) ways. Unfortunately, the benefits, risks and harms for that technology are not equally distributed. I'm concerned about the resulting dissonances. In particular, I'm concerned about those dissonances, as all machine learning tools have one common denominator: they are a statistical technology. Statements about individuals are derived from measurement about aggregates. Your job performance is derived from the job performance of your similar peers. The next token in a series of language tokes is derived from a multitude of follow-up tokens observed in a large corpus of text. Statistical technologies, however, are a rather modern invention (in comparison to such technologies as electronical ones). Its ubiquitous modern use demands to ask some of the fundamental and basic questions, potential again. We need to understand it better, to mitigate the harmful dissonances and strengthen the constructive benefits.

Currently, I'm working on "theories of practices of data" :) That's a mouthful, but captures essentially my constructive approach to fill the gap of limited theorization of data construction, despite data's prominence in machine learning pathways. I want to use tools of (and always interested in) game-theoretic probability and randomness, hypergraph theory and describe data construction phenomena such as changes in measurement in concrete empirically interesting cases, e.g., in the quantitative social sciences or clinical setups. I hope to bridge disciplinary boundaries as well as formal and empirical thinking.

I previously was a PhD Student in the same FMLS lab. From October 2023 until March 2024 I did a research stay at the University of Pennsylvania visiting Aaron Roth. The paper which resulted from the stay is a this great discovery tour into game theory -- at least for me :) The Value of Ambiguous Commitment in Multi-Follower Games . I was a scholar in the International Max Planck Research School for Intelligent Systems (IMPRS-IS) from 2021 to 2026. I defended my dissertation "Four Facets of Forecast Felicity" in May 2026, in which I elaborate on measurability and randomness as relativized assumptions required for statistical forecasts, and I explore semantics and structure of how such forecasts are evaluated.

I'm a co-founder of Catalyst and a strong believer in the constructive support and cooperation within the scientific community. For the same reason, I am an active member of the Tübingen AI PhD PostDoc Assembly (TAPPA).

I love reading books and just started to create "ordered list of alphabetical symbols with semantics" myself. My short-story series "Half a Story" on machine learning just started with:

Neosome and the Prophet or How Does Machine Learning See Me? (see below)

To be continued...

The State of the Art or How Does Machine Learning Make Progress?

The Median Man

You can see me giving talks at...

Tübingen AI Center: https://www.youtube.com/watch?v=scQxHRVC0Cw&pp=ygUMcmFiYW51cyBkZXJy on "The Value of Ambiguous Commitments"

CISPA (Saarbrücken): https://www.youtube.com/watch?v=Q-czMvfN79M on "Law of Large Numbers: Accuracy as Statistical Measure for Competition and Compliance"

For a list of publication see the file below.

Contact

University of Tübingen
Department of Computer Science
Maria von Linden Str. 1
72076 Tübingen
Germany

Room: A-433
Phone: +49 (0)7071 29-70840
Mail: rabanus.derr[at]uni-tuebingen.de