Project Reproducibility and Simulation avoidance
Science, especially scientific computer simulations aim to be reproducible. HPC simulations as for example done here at HLRS often run into reproducibility issues, caused not only by infrastructural issues such as dark data or backward compatibility. Because of parallelization it can be the case that even bit-wise reproducibility is not achievable. This poses somewhat of a dilemma for simulation science. There is an expectation of strong reproducibility which can in practice only partially be fulfilled. One goal of this project is to shed light on the notion of reproducibility in simulation science, comparing it to similar notions from neighboring fields such as computational statistics. How should reproducibility expectations be balanced with what simulations can realistically deliver? Another question which I try to answer is what kind of reproducibility we should expect from hybrid and surrogate simulations. Currently ML methods, because of their cheaper computational costs are introduced in the simulation pipeline everywhere. They come with their own reproducibility issues, stemming from different sources than those in computer simulations. I try to figure out how these issues interfere and what kind of properties computational problems must have to be amenable to the “ML-treatment”.
Blog
You can find scattered thoughts mostly on the philosophy of computational methods on the workshopping blog I run with Ramón Alvarado. See also my new Mastodon account: Mastodon
Scientific interests
- Philosophy of science/philosophy of technology
- Intertheoretical relations in physical theories (reduction, derivation, founding, modularity etc.)
- Computer simulatable physical theories (e.g. LatticeQCD)
- Computational reliabilism
- Constructing artifacts as pragmatic justification of scientific theories
- discrete vs continuous science
- Limits of statistical inference/learning
- Inductive assumptions in approxmative methods
- Machine Learning & the problem of induction
- Kolmogorov complexity as general framework for ML
- theoretical vs. practical limits (e.g. no free lunch theorems vs. FLOPs)
- Science and society
- Why trust science?
- Why trust society?
Work In Progress
Working for trust – when easy things become hard (Abstract)
The theory and practice of computational error (Abstract – of a talk with the same title given at IACAP 23)
Recent and forthcoming publications
- A minimalistic definition of XAI explanations (with Colin Glass), Springer, 2025
- Can computer technology change physical theories? Jahrbuch Wissenschaftsforschung 2023, WVB, 2025
- Inductive assumptions in ML – often ignored, always required, Springer, 2025
- How I stopped worrying and learned to love opacity, Chapter for Philosophy of science for machine learning: Core issues and new perspectives (Juan Manuel Durán & Giorgia Pozzi (eds.)), Springer, forthcoming 2025?
Publications of note:
- Durán, J.M., Formanek, N. Grounds for Trust: Essential Epistemic Opacity and Computational Reliabilism. Minds & Machines 28, 645–666 (2018). https://doi.org/10.1007/s11023-018-9481-6
Upcoming Talks
- The role of computation in theory choice (Kolloquium Aachen/Darmstadt/Stuttgart, January 22nd, 26)
Recent talks
Slides in links.
- Can we automate scientific decisions? (Workshop on automated science, U Oregon, November 24th, 25)
- Computing power and model selection (OCIE, Chapman University, November 21st, 25)
- Ethics in HPC BoF (SC25, St. Louis, November 18th, 25)
- Computer haben keine Angst vor Fehlern (Symposium Vom Fehler und seinem Potential, Linienscharen, November 8th, 25)
- Computing power and model selection (Wissenschaftstheoretisches Kolloquium, Uni Bern, November 4th, 25)
- Reproducing Simulations (IACAP25, Twente, July 2nd, 25)
- How could a HPC center ever be sustainable? (PASC, Brugg-Windisch, June 17th, 25)
- Power and Performance (WSSP, Stuttgart, May 28th, 25)
- Generalization (or Overfitting) (What is a good model?, Stuttgart May 27th, 25)
- Are hyperparameters vibes? (IRIS Insights, April 24th, 25)
Teaching (Current and Past Semester)
- Philosophy of Machine Learning (Uni Stuttgart) Syllabus – Literature List
- Philosophy of Computer Science (Uni Stuttgart) Syllabus – Literature List
Simulierte Welten
Current Project:
Currently (25/26) I am mentoring two high-school students for Simulierte Welten on the topic of analog computing. We are building simple analog computers on a bread board to understand how they operate and what kind of calculations they can do. In the end we hope the compare an analog implementation of Lorenz’ 63 model of turbulence (the model is famous for being involved in the discovery of deterministic chaos) to its common digital implementation.
Past Projects:
In 24/25 I supervised three high-school students in the project Simulierte Welten. We are implemented a simplified version of Langley’s BACON program for automated scientific inference on the arduino platform. The goal was to use the on board sensors of the arduino nano sense to automatically discover simple physical laws and get an idea about the robustness of the BACON method to noise.
In 23/24 I supervised two high-school students in the project Simulierte Welten. We developed a topical model for the PhilosL mailinglist. This model automatically extract topics (i.e. metaphysics, philosophy of science, ethics) from the archives of the mailing list, allows to observe trends in topical distribution and evolution – and correlate topics with other interesting quantities. A more detailed description of the project can be found here (in German).
