Komplexitätsreduktion, Erklärbarkeit und Interpretierbarkeit (KEI)
-
- Junge Akademie
- WIN-Forschungsprojekte
- Komplexitätsreduktion
- Komplexitätsreduktion, Erklärbarkeit und Interpretierbarkeit (KEI)
Auf der Suche nach erklärbarem und interpretierbarem Maschinellen Lernen mit Philosophie und Physik
Algorithmen des Maschinellen Lernens (ML) durchdringen zunehmend unseren Alltag und das öffentliche Leben. Sie treffen Vorhersagen, aber warum sie so und nicht anders entscheiden, bleibt oft schwer nachvollziehbar, sie sind gewissermaßen „opak“. In unserem Projekt wollen wir verstehen, wie diese Opakheit entsteht und wie man sie rückwirkend aufheben könnte. Dafür wollen wir anhand von Einsichten der Physik und anderer Theorien der Komplexität die Natur der (impliziten) Abstraktionen interpretieren, die ML an sich erzeugt. Unsere Arbeitshypothese ist, dass die Komplexität des ML und die Schwierigkeit, gewisse Komponenten des Lernprozesses zu verstehen, gemeinsam das Problem der Opakheit hervorbringen. In diesem Sinne fordert eine Lösung nicht einfach „mehr Verständnis”, oder „weniger Komplexität”, sondern eine sinngebende Komplexitätsreduktion. Damit meinen wir adäquate Abstraktionen und nicht triviale Vereinfachungen, die einen wohl fundierten Verständniszugang gewährleisten. In unserem Projekt werden wir Werkzeuge entwickeln, um die Komplexität der ML-Algorithmen auf neue Weise zu analysieren und sinngebende Reduktionen aus der Perspektive der Vielteilchenphysik und der Philosophie aufzufinden.
WIN-Kollegiatinnen und -Kollegiaten
Publikationen peer-reviewed
Bordt, S.; Raidl, E.; von Luxburg, U. (2025): Position: Rethinking Explainable Machine Learning as Applied Statistics, Proceedings of the 42nd International Conference on Machine Learning, PMLR 267: 81130–81142.
Publikationen unter peer-review
Klopotek, M. (2026a): Fluctuations in Dynamical Environments: Redefining Computation Emerging, Book chapter for Yearbook for Philosophy of Complex Systems (2nd Ed.) [Eds.: Fraisopi, F. and Saratxaga, Arantzazu].
Raidl, E. (2026): Logics and semantics for “because”, Journal of Logic Language and Information, (2nd revision).
Publikationen nicht peer-reviewed
Wetzel, S., … , Klopotek, M. et al. (2025). Interpretable Machine Learning in Physics: A Review, arXiv:2503.23616.
Stein, J. and Raidl, E. (2026): How Complexity Contributes to Learning Opacity in Machine Learning, pp. 1-32, arXiv:2606.24953
Weinmann, M. and Klopotek, M. (2026a): Interpreting learning dynamics of autoencoders: Transient scaling and emerging concepts of the Ising model, pp. 1-59, arXiv preprint arxiv:2607.10285.
Organisierte Tagungen
Research Frontiers Workshop on Scientific Machine Learning: Navigating the Bermuda Triangle of Knowledge Infusion, Explainability, and Scientific Discovery (A. Guthke, M. Klopotek, E. Raidl, A. Totounferoush) (07-10.10.2025, Stuttgart). Partly funded by the HAdW/KEI project.
Organoid Intelligence Workshop (A. Wernick, M. Klopotek & D. Kronenberg-Versteeg), Stem Cells in Neuroscience Conference (26.02.2026, Tübingen).
Vorträge (inkl. Reden, zus. Poster)
Bordt, S.*; Raidl, E.; von Luxburg, U. (2025): Position: Rethinking Explainable Machine Learning as Applied Statistics (42nd International Conference on Machine Learning, ICML 2025, Vancouver, 13-19.7.2025).
Egenlauf, P*., Březinová, I., Andergassen, S., and Klopotek, M. (2026b): Neural ODEs for Reduced-Order Quantum Many-Body Dynamics: Assessing Memory Effects (32nd Meeting of the Condensed Matter Division, European Physical Society (CMD) 20-25 September 2026, Graz Center of Physics, Austria).
Egenlauf, P*., Březinová, I., Andergassen, S., and Klopotek, M. (2026b) [poster]: Neural ODEs for Reduced-Order Quantum Many-Body Dynamics: Assessing Memory Effects (Roccella Conference on Inference and AI - ROCKIN' AI, 31.08- 5.09 2026, Roccella Jonica, Italy)
Egenlauf, P*., Kröninger, H., Kung, A., and Klopotek, M. (2026b): From Phase Space Fluctuations to Predictive Power: Entropy Production as a Metric for Swarm Reservoir Computing (Spring Meetings of the German Physical Society (DPG): Condensed Matter Section 8-14 March 2026, 09.03.2026, Dresden).
Gaimann, M. U. and Klopotek, M.* (2026c): Performing Inference with Physical Response: Reservoir Computing with Active Matter Substrates“ (Spring Meetings of the German Physical Society (DPG): Condensed Matter Section 8-14 March 2026, 09.03.2026, Dresden).
Gaimann, M. U.* and Klopotek, M. (2026d): Reservoir Computing with Active Matter Systems (APS Global Physics Summit, American Physical Society, 18.03.2026, Denver, CO, USA).
Klopotek, M. (2024): A reflection on computational modeling… (SAS24 Conf. on Modeling for Policy, HLRS Stuttgart, 11/2024)
Klopotek, M. (2025a): Statistical Physics and Machine Learning (given to scholarship holders of the Hans-Böckler-Stiftung, 18.03.2025, Stuttgart.)
Klopotek, M. (2025b): From ML interpretability to robustly optimal information processing with active-matter reservoir computers (colloquium at the Technical University of Vienna / invited via I. Brezinova and S. Andergassen, 35.06.2025, Vienna.
Klopotek, M. (2025c) [speech]: The cloud of unknowing: A journey towards AI with physics (commencement/keynote at the Central Doctoral Graduation Ceremony of the University of Tübingen, 19.07.2025, Tübingen).
Klopotek, M. (2025d): Physical roots of computation and learning in malleable uncertainties; Do we need a physical definition of inference? (SAS25 Conference - Uncertainty, HLRS Stuttgart, 07/2025)
Klopotek, M. (2025e): Complexity (Research Frontiers Workshop on Scientific Machine Learning: 6-10.10.2025, Stuttgart)
Klopotek, M. (2026b): Reservoir Computing with Biological-like Matter: Basic Perspectives for Future OI [Organoid Intelligence] (Organoid Intelligence Workshop, Stem Cells in Neuroscience Conference, 26.02.2026 Tübingen).
Klopotek, M. (2026c): Basic physics insights from statistical mechanics towards future computing (Future Computing Workshop, HLRS, 26.03.2026, Stuttgart).
Klopotek, M. (2026d): Basic physics insights for embodied intelligence from models of computing in materio (Embodied Intelligence Conference, 20.03.2026, global). Proceedings online [https://www.youtube.com@EmbodiedIntelligenceConference].
Klopotek, M. (2026e): Reservoir Computing with Active Matter: A Statistical-Physical Viewpoint (International Reservoir Computing Conference, 25-27 March 2026, 25.03.2026, Berlin). Proceedings recorded.
Klopotek, M. (2026f): Physical reservoir computing with active matter: Fundamental insights into learning and inference as non-equilibrium statistical mechanics (Academy Day of the Johanna Quandt Young Academy), 08.05.2026, Frankfurt).
Klopotek, M. (2026g) [speech]: Bericht der Jungen Akademie (Annual Celebration of the Heidelberg Academy of Sciences and Humanities, 20.06.2026, Heidelberg).
Klopotek, M.* & Raidl, E.* (2026): Maschinelles Lernen Verstehen – eine physikalische Perspektive (Vortragsreihe Wir forschen für Sie, 22.7.2026, Heidelberg).
Raidl, E. (2025): Explainability (Research Frontiers Workshop on Scientific Machine Learning: 6-10.10.2025, Stuttgart).
Raidl, E. (2025): Logics for “Because” (Workshop zum “Because”, 28.11.2025, Mannheim).
Raidl, E. (2026): Explainability and Complexity (Values in Machine Learning, 16-17.10.2026 Konstanz).
Stein, J. (2025): Statistical Learning Theory meets Formal Learning Theory - Ockham's Razor and VC Dimension. (German Society for Philosophy of Science Conference, Erlangen, 24.-26.03.2025).
Stein, J. (2026): Counteracting Complexity Driven Opacity in Machine Learning. (University of Groningen TF-PCCP Meeting, Groningen, 04/2026).
Stein, J. (2026): Does Mechanistic Interpretability Research Provide Mechanistic Explanations? (SOPhiA 2026 - Salzburg Conference for Young Analytic Philosophy, Salzburg, 02.-04.09.2026).
Stein, J.* & Raidl, E. (2025a): On the Complexity of Neural Networks. (Tübingen and Friends Workshop for Philosophy of Machine Learning 2025, Tübingen, 04/2025).
Stein, J.* & Raidl, E. (2025b): Complexity and the Opacity of Neural Network Training. (SAS25 Conference on Uncertainty, HLRS Stuttgart, 28.-30.07.2025).
Weinmann, M.* and Klopotek, M. (2024): Contextual Alignment for Robust Learning in Dynamical Systems (Spring Meetings of the German Physical Society (DPG): Condensed Matter Section, 03/2025, Berlin).
Weinmann, M.* and Klopotek, M. (2026a): Autoencoder Learning Dynamics on MCMC Ising Dataset (Spring Meetings of the German Physical Society (DPG): Condensed Matter Section 8-14 March 2026, 09.03.2026, Dresden).
Weinmann, M.* and Klopotek, M. (2026b) [poster]: Learning Dynamics of Autoencoders on Ising model data (Roccella Conference on Inference and AI - ROCKIN' AI, 31.08- 5.09 2026, Roccella Jonica, Italy).
Weinmann, M.* and Klopotek, M. (2026c): Learning Dynamics of Autoencoders on Ising model data (32nd Meeting of the Condensed Matter Division, European Physical Society (CMD) 20-25 September 2026, Graz Center of Physics, Austria).
Weinmann, M.* and Klopotek, M. (2026d) [poster]: Learning Dynamics of Autoencoders on Ising model data (2026 Bootcamp of the IMPRS-IS, 23-25 Sep 2026, Sonthofen).
(* indication of the presenter of a contribution with multiple authors listed)
Veranstaltungen o. Publikationen in der Wissenschaftskommunikation/ Transfer
Raidl, E. (2026): Interview mit André Boße für das Magazin „Perspektiven“ der Baden Württemberg-Stiftung. [André Boße: „Soll ich dir sagen, wie’s dir geht?“ Über KI und Kommunikation. In: Perspektiven - Das Magazin der Baden-Württemberg Stiftung. Ausgabe 01/2026.]
Klopotek, M. (2026d): Prometheus: Von der physikalischen Emergenz zur Diversifizierung der Künstlichen Intelligenz und des Rechnens, Athene: Magazin der HAdW (June 2026 Ed.).
Sonstige Projektrelevante Publikationen der Antragstellenden
Egenlauf, P., …, Klopotek, M. (2026a) Capturing reduced-order quantum many-body dynamics out of equilibrium via neural ordinary differential equations, Machine Learning: Science and Technology 7(2), 025062.
Egenlauf, P., …, Klopotek, M. (2026b) Entropy production of active matter systems as an indicator of computing performance, to appear on arXiv.
Gaimann, M. U. and Klopotek, M. (2026a): Reservoir computing with active matter, Part I: Robustly optimal intrinsic dynamics. Under Review (final round), arXiv:2505.05420 preprint (01/2026, v4).
Gaimann, M. U. and Klopotek, M. (2026b): Reservoir computing with active matter, Part II: Optimal information injection and propagation mechanisms. Under Review, arXiv:2509.01799 (3/2026, v2).
Gaimann, M. U., Romero Castillo, A. and Klopotek, M. (2026): Static-Shape Input Encodings as Baselines for Active Matter Reservoir Computing. Manuscript, to appear on arXiv.
Klopotek, M. (2026a): Relations of adaptivity, robustness, and interpretability in beta-VAEs… (working title).
Klopotek, M. (2026h): Nonequilibrium physics of machine inference in active matter reservoir computers (working title)m [Invited article].
PD Dr. Eric Raidl
Exzellenzcluster "Maschinelles Lernen: Neue Perspektiven für die Wissenschaft"
AI Research Building
Maria-von-Linden-Str. 6
72076 Tübingen
Dr. Miriam Klopotek
Universität Stuttgart
Stuttgart Center for Simulation Science
SimTech Cluster of Excellence
Universitätsstraße 32
70569 Stuttgart