Representative Signatures in Data Analysis - Epistemological Reflections

15 Oct 2026, 10:00 → 16 Oct 2026, 18:00 Europe/Berlin

JvF25/3-303 - Conference Room (Lamarr/RC Trust Dortmund)

The workshop "Representative Signatures in Data Analysis - Epistemological Reflections" will bring together philosophers of science and colleagues from astroparticle physics, computer science, and statistics. It will focus on the way in which machine learning models manage to extract significant data from data sets that contain a huge background of irrelevant data. Current methods of statistical data analysis base the extraction of the significant data on probabilistic quality  standards for representative signatures and employs machine learning models to extract the relevant data subsamples.

At the workshop, we want to address the following questions:

• How have the current machine learning methods transformed the analysis of observational data, compared to the scanning of data samples by human beings, for example in particle physics from the 1960s to the 1980s?
• To which kind of representation do the probabilistic quality standards implemented in the AI models give rise?
• In which sense and to what extent result these methods in representing real data or true events?
• Which role plays the "inverse problem" of tracing the measured raw data back to event distributions considered to be approximately true, from a philosophical point of view?
• To what extent is the representation of data by AI models compatible with scientific realism about the results of the data processing?

The workshop will take place on October 15-16 at TU Dortmund University.

Thursday 15 October

Introduction

Unfolding

1 Knowledge, Big Data, and Artificial Intelligence - From Critical to Probabilistic Rationalism

This talk explores the shift from classical critical rationalism to probabilistic rationalism in modern experimental physics. Using Plato's Cave as a metaphor, it argues that modern detectors produce probabilistically smeared shadows of underlying physical processes, making the inference from data to theory inherently non-unique. The presentation identifies three epistemological shifts in modern experiments: high-dimensional data volumes, the granularity of raw measurements, and the role of prior knowledge encoded as Bayesian priors. Because the inverse problem is ill-posed, regularization and controlled assumptions are required, and the resulting bias must be transparently quantified. Monte Carlo simulation provides a virtual reality for training machine-learning models, while Bayesian inference updates the plausibility of hypotheses rather than proving them.
The talk concludes that modern knowledge acquisition is a cyclical process in which AI extends rather than replaces scientific rationality. The resulting framework holds that knowledge remains rational and critical, yet fundamentally probabilistic: what is obtained is not a direct observation-to-theory inference, but a probabilistic statement grounded in data, model, simulation, and inference.

Speaker: Prof. Wolfgang Rhode (TU Dortmund)

2 Unfolding data with AI at the LHC: Epistemological implications

Unfolding is an inverse data transformation process that aims to correct high-level physics observable quantities obtained from a set of data for possible distortions introduced by the instrument. Unfolding is a crucial step to enable experimental outcomes to be directly comparable to theoretical predictions. Traditional unfolding methods used in experimental High Energy Physics consist in statistical inference of a particle-level distribution from a corresponding detector-level distribution. While capable of high accuracy and precision, these techniques have many shortcomings including limitation in scalability, flexibility, and dependence on simulations. Here we discuss a novel AI-based approach to multidimensional object-wise unfolding, first introduced by [Pazos, Beauchemin et al] using conditional Denoising Diffusion Probabilistic Models (cDDPM). While addressing many of the challenges posed by traditional unfolding methods, this new approach offers the opportunity for interesting philosophical insights. For example, this approach blurs the distinction between low-level pattern identification and high-level information about physics processes. Another epistemologically interesting aspect of this algorithm is that it requires inductive bias to better approximate a universal unfolding tool, making theory-ladenness a desideratum. Finally, the AI-based algorithm does not seem more or less opaque than the traditional statistical-based approach, highlighting the importance of the critical assessment of the performance of any algorithm, and indicating a continuity between machine-learning methods and previous statistical methods for the analysis of observational data.

Speaker: Prof. Pierre-Hugues Beauchemin (Department of Physics and Astronomy, Tufts University, Department of Philosophy)

3 Unfolding, Systematics, and Trustworthiness

Modern unfolding techniques pay special attention to
systematic effects, such as imprecisely known or varying properties of the detection process. This necessary, although fairly novel, shift towards the modeling of systematic effects corresponds to a desirable relaxation of assumptions previously imposed by the physicist, so that unfolding can represent the observed data more closely than historical unfolding techniques could. We recently studied this relaxation from a learning-theoretic viewpoint, finding that it demands more computational resources and closer attention from the physicist. Specifically, this relaxation yields only local consistency of unfolding techniques, whereas the historical assumption yields global consistency. These two levels of statistical consistency correspond to two different levels of trustworthiness: methods enjoying global consistency are trustworthy as soon as the (inappropriately strict) historical assumption holds, but not otherwise; methods with only local consistency instead require the physicist to properly handle multi-start optimization, or other heuristics for global non-convex optimization.

Speaker: Dr Mirko Bunse (Lamarr Institute, TU Dortmund University)

12:25 Lunch break

Signal-background separation and data quality

4 Statistical separation of signal and background based on partial prior knowledge

Data analysis is often faced with the problem to determine the properties of a signal in the presence of background. A measurement is possible if the two components can be statistically separated in a discriminant variable, provided the probability density functions in this variable are known or can be estimated from the data. This is realized by Custom Orthogonal Weight functions (COWs). The talk explains the method and illustrates it by means of numerical examples.

Speaker: Prof. Michael Schmelling (Max-Planck-Institut für Kernphysik Heidelberg)

5 The impact of data quality and representativity for statistical learning models

The quality of results of machine learning models highly depends on the quality of the data used to train the model. But what do we mean by good data quality? Are representative data, e.g., provided by the German micro census or the census of good quality? And does the term “representativity” suffice to define data quality? In this talk, we will discuss the terms “data quality” and “representativity” and link them to the learning task of statistical models.

Speaker: Prof. Katja Ickstadt (TU Dortmund)

15:30 Coffee break

Signatures in statistical data of (astro-)particle physics

6 Data flows in observational astrophysics

Research in astronomy and astrophysics has over centuries made huge progress along the frontier of ever increasing observational capabilities, which supply a wealth of data spanning the electromagnetic spectrum and different messengers.
With this treasure drove of data sitting in archives and being continuously added to, questions remain on how to best ensure equal access for members of the global community, and the availability of data for an indefinite timeframe. This is made even more challenging by the
fact that some observational capabilities exist only in few places in the world, and that some astronomical events (well known historic examples would be Galactic supernovae) are so rare that they will not
repeat during a single human lifetime.
In this talk, I will give an overview on challenges connected to this situation, and on efforts to ensure lasting data availability and access. In addition, I aim to discuss the potential use cases for artificial intelligence in this context.

Speaker: Dr Dominik Elsässer (TU Dortmund)

7 Signatures of electrons at ATLAS: The entheorization of data

The study of particle signatures at High Energy Colliders provides a concrete context to think about pattern identification in experimental data. An inquiry into pattern identification in relation to particle signatures stands to yield new insights on long-standing philosophical questions about patterns (highlighted in an influential paper by Dennett), such as their relationship to phenomena (Bogen and Woodward, Feest), or their dependence on theory and on the epistemic objectives of inquiries (McAllister). Discussions about the relationship between data and data models (Suppes, Harris, Leonelli) also provide interesting perspectives on the evidential role of patterns in experimental knowledge. In this presentation, we use particle signatures, such as that of the electron, to reflect on the notion of pattern. We analyze in detail the data analyses performed by the ATLAS Collaboration at CERN that serve to identify electron signatures, to support a new framing of patterns in pragmatist terms. We have used a pragmatist framework previously to address important philosophical problems related to exploratory experimentation, the usefulness of measurements, and measurement uncertainty and its relation to underdetermination. Our account integrates some of the important points about patterns made by or inferred from the work of Harris, McAllister, and Leonelli, but it addresses some gaps in and shortcomings of their views while also differing from the approach taken by Mättig and Stöltzner in their work on signatures. We provide an account of patterns as resources to be used in an inquiry to yield useful scientific claims. Their usefulness in inquiry depends on a transformative process of “entheorization” that intertwines theory and experimental practices together. It also depends on decisions, made in light of experimental objectives, about tradeoffs between signal purity and efficiency, and regarding acceptable uncertainties. The assertability of conclusions reached through these aim-guided processes rests upon their being subjected to a critical mode of inquiry.

The picture that emerges from this study of the relationship between data, patterns in data, and particle signatures, represents a challenge to a view of signatures as providing a stable and theory-agnostic representation of electrons or their experimental behavior (Mättig & Stöltzner). Rather, signatures constitute a means – a resource -- for entheorizing data into useful patterns that in turn support experimental objectives.

Speaker: Prof. Kent W. Staley (Department of Philosophy, Saint Louis University)

17:15 Coffee break

Signatures in statistical data of (astro-)particle physics

8 A generalized concept of signatures?

In a recent paper, Peter Mättig and I developed a generalized version of the concept of signatures broadly used in elementary particle physics and discussed its relation to phenomena and data models. My paper explores, for one, how this concept can be extended to other fields and where such an extension is most likely to fail. Secondly, I discuss to what extent signatures can be a target for AI-assisted searches for new physics and whether modifications of this concept are necessary.

Speaker: Prof. Michael Stöltzner (Department of Philosophy, University of South Carolina)

9 Tracing Particles: “Scanning Girls” and the Automation of Image Evaluation at the German Electron Synchrotron (DESY)

In the course of the “industrialization” (Galison 1997) of high-energy physics during the 1950s and beyond, the evaluation of vast quantities of experimentally produced photographs was initially delegated to untrained women. However, their “practices of seeing” (Schürmann 2008) were soon identified as the economic and epistemic bottleneck of the experimental process — a barrier that was to be overcome through computerized automation.
Internationally, as well as at the Hamburg accelerator (DESY) in West Germany, significant potential was seen in the flying spot scanner. Within this media network, the so-called Hough Powell Device (HPD) and its associated computer systems were designed to take over complex pattern recognition tasks, effectively automating the extraction of significant signatures from the data background.
Although this “digital” system—viewed by Peter Galison as a precursor to Artificial Intelligence — aimed to standardize data analysis, it remained deeply entangled with local, gendered practices and specific conditions of knowledge production. Based on archival research, oral history interviews, and re-enactments, this paper investigates the epistemic, practical, and technical dimensions of these local knowledge practices within the “HPD logic” (Lingjaerde 1962). It specifically explores how the shift toward automated pattern recognition reconfigured gendered labor and challenged traditional notions of scientific objectivity in the production of experimental data.

Speaker: Dr Dinah Pfau (Institute for the History of Science and Technology, Deutsches Museum Munich)

20:00 Dinner

Friday 16 October

Models and representation in astroparticle physics

10 Models and Signatures in Astroparticle Physics

Astroparticle physics measures cosmic rays using particle detector systems that function as telescopes for detecting subatomic particles of extraterrestrial origin. As part of current multi-messenger astronomy, its models and methods bridge the gap between astrophysics, cosmology, and particle physics. My talk focuses on the causal models underlying the data analysis of the IceCube neutrino observatory at the South Pole and the probabilistic features of these models. Based on recent IceCube results, my talk will focus on three questions: (1) What kind of causal modelling underlies the probabilistic data analysis? (2) How does this causal modelling relate to philosophical accounts of causality and probability? (3) What kinds of signature are crucial for the identification of specific neutrino events and the causal reconstruction of their origin?

Speaker: Prof. Brigitte Falkenburg (TU Dortmund)

11 Scientific Models as Representations: From Astroparticle Physics to Physics Education

Modern physics is deeply shaped by the development and use of models. From detector simulations and statistical inference to the interpretation of experimental data, scientific knowledge is rarely derived from observations alone but emerges through layers of modeling that represent physical processes, instruments, and uncertainties.

At the same time, the role of models as representations of the physical world often remains implicit in how physics is taught in schools. In school contexts, physical laws and formulas can appear as direct descriptions of reality rather than as idealized representations with specific assumptions, limited domains of validity and associated uncertainties. As a consequence, learners frequently encounter difficulties in revising or replacing previous models when more refined ones are introduced. More broadly, this also limits their general understanding of the provisional character of scientific knowledge and the role of uncertainties in scientific reasoning.

Drawing on examples from astroparticle physics and perspectives from physics education research, this talk examines models as epistemic tools that mediate between phenomena, experiments, and observable data. Making this model‑based structure of physics explicit - both in research practice and in education - can help bridge the perceived divide between “school physics” and “real physics”. Developing an understanding of models as representations is therefore not only central to learning physics, but also increasingly important for navigating a world shaped by scientific modeling, data analysis, and algorithmic methods.

Speaker: Dr Marlene Doert (Department of Physics, TU Dortmund)

11:30 Coffee break

Deep learning and scientific realism

12 The Prospects for Realism in AI-Driven Science: The Case of Exoplanet Detection

In this talk, I ask whether and how the use of machine learning (ML) in contemporary science affects the scientific realism debate. After discussing the current state of the debate, I focus on the case of exoplanet detection, where ML has recently been used to distinguish transit signals produced by planets from false positives. I ask, first, whether ML can provide a methodologically distinct route that converges with established statistical approaches such as vespa, thereby strengthening the no-coincidence argument for realism; and second, whether ML-based anomaly detection can make exoplanet science less vulnerable to the problem of unconceived alternatives by identifying cases that fall outside established categories. The responses to these questions point toward a cautious optimism about the prospects for realism.

Speaker: Dr Mahdi Khalili (Institute of Philosophy, University of Bern)

12:30 Lunch break

Deep learning and scientific realism

13 What can we learn from hallucination? Scientific realism and models in astrophysics in the age of the AI

In recent work, Constantin et al. (2026) suggested the development of language models that can generate coherent mathematical representations, advancing the automation of physical law discovery. Thanks to structural insights, they believe that one can inform the training or fine-tuning of language models for symbolic reasoning, encouraging them to generate expressions whose statistical and structural features mirror those found in real physics. Such an approach seems to be highly compatible with scientific realism, the “positive epistemic attitude toward the content of our best theories and models, recommending belief in both observable and unobservable aspects of the world described by the sciences” (Chakravartty 2011). We, however, know that LLMs are susceptible by necessity to hallucination and that it is impossible to avoid it. Interestingly, hallucination impacts AI-generated models in similar ways as systematic and random errors affect experimental physics and astrophysics. After exploring this analogy, I suggest an interpretation of hallucination that allows one to consider scientific realism (in a constrained form) compatible with the representation of data obtained through AI models. Finally, by discussing recent examples in agentic AI for science, and astrophysics in particular, I show that we need human understanding and interpretation of mathematical models to identify epistemic structures and practices in astrophysics in a meaningful way.

Speaker: Dr Silvia De Bianchi (Department of Philosophy, Università degli Studi di Milano)

14 The Explanation–Justification Gap in Deep Learning Astroparticle Physics

Deep learning (DL) models are increasingly used in astroparticle physics for tasks such as gamma–hadron separation, neutrino event reconstruction, and cosmic-ray classification. While these models achieve remarkable predictive accuracy, their opacity poses a challenge to the epistemic standards of discovery. Heatmap-based explainable AI (XAI) techniques —such as heatmaps — promise insight into model reasoning, yet visualization alone cannot justify scientific claims. This paper identifies the explanation–justification gap and proposes epistemic preconditions for closing it. By situating these conditions within contemporary practices of detector-based inference, the paper clarifies when heatmaps contribute to justified knowledge in astroparticle physics.

Speaker: Dr Koray Karaca (Middle East Technical University Ankara)

15:30 Coffee break

Deep learning and scientific realism

15 Does machine-learning-driven scientific discovery pose new challenges to the No Miracles Argument?

The No Miracles Argument is arguably the strongest case for scientific realism. Yet, it seems that machine-learning-driven scientific discovery evades its justificatory force. Machine learning models are explicitly trained for the purpose of detecting patterns in data, their predictive success comes by design and does not indicate the truth of the underlying representation. Moreover, such models do not generalize well beyond the domain of their specific training data. Thus, there seems to be no miraculous success to be explained.

However, I will argue that this line of reasoning does not apply to the machine learning models that are currently employed in physics research. Due to their relatively restricted tasks such as signal-background separation, solving the inverse problem, or hypothesis tests, the crucial inferential steps remain in human hands. Such machine learning methods might be faster and more accurate but not essentially different from previous methods.

Speaker: Dr Johannes Mierau (TU Dortmund)