The talk examines how astroparticle physics (APP) combines machine learning with causal-mechanistic explanations. To separate data from background in cosmic ray measurements, machine learning is employed. This separation results in probabilities for attributing signals to cosmic sources. APP combines these probabilistic results with causal stories about messenger particles that transfer information from cosmic sources to the Earth, raising philosophical problems regarding the theory-ladenness of data and their interpretation in terms of scientific realism vs. instrumentalism. For clarification, a closer look at the data analysis of APP is taken here, focusing on two questions: How does the machine learning used in the data analysis model the physical data-background-relation? How can probabilistic data give rise to causal stories that explain single signals? As a case study, the neutrino signal measured in 2017 by the IceCube detector in Antarctica serves. I argue that here machine learning was a mere statistical tool, whereas the causal interpretation tracing the signal back to a cosmic source depended crucially on a coincidence with the high-energy gamma ray activity of the blazar TXS 0506+056, indicating the significance of independent evidence and common-cause-arguments in multi-messenger astronomy.

Several authors have recently made a connection between algorithmic bias in machine learning and the debate about the value-free ideal in the philosophy of science. Using arguments that set out from the underdetermination of inductive conclusions by the data, a trait shared by scientific reasoning and machine learning algorithms, some have concluded that machine learning algorithms *must* be value-laden.

Drawing from the philosophy of induction, I aim to show why a general argument from underdetermination does *not* already settle that learning algorithms are value-laden. I first clarify the relevant notions of “value-ladenness” and “machine learning algorithm”. Then, relying on a distinction between domain-general learning rules and their domain-specific inductive biases, I argue that underdetermination does not yet preclude that either component is epistemically motivated or even justified. I illustrate my observations by the general framework of Bayesian induction and by the specific example of a classical convolutional neural network.

Deep Neural Networks (DNNs) are transforming fields like Particle Physics, but their black box nature challenges scientific understanding. This talk explores whether DNNs can serve as vehicles for scientific understanding and what conditions must be met for this to occur.

I propose a minimal account of scientific understanding distinguishing between understanding as a state and a process. Scientific understanding requires both (moderate) factivity and usability. Additionally, representational devices must be intelligible and allow reliable inferences to provide understanding. The problem is that DNNs themselves are often not understood, making it difficult to use them as vehicles for scientific insight.

A promising direction lies in eXplainable AI (XA)), but many methods focus on how DNNs work rather than what they reveal about physics. To regain understanding, researchers must extract meaningful structure from DNNs, not just interpret their internal mechanics. However, a fundamental dilemma remains: intelligiblee methods often solely confirm known physics, while novel insights frequently lack clear scientific interpretation. I argue that bridging this gap is crucial for making DNNs genuine tools for scientific understanding and will discuss ongoing efforts in Particle Physics to tackle this challenge.

In this paper, I discuss the ambiguity problem which concerns the heterodox view on the qualitative identity/difference of similar fermions. According to the heterodox view, similar fermions are (almost always) qualitatively distinct, so that Leibniz’s Principle of the Identity of Indiscernibles is valid (this view contrasts with the orthodox view, according to which the opposite is the case). However, several authors have pointed out that there is an ambiguity problem for the qualitative individuation of similar fermions, i.e., individuation via properties is not unique. More specifically, there is a type of ambiguity according to which fermions sometimes possess definite spin properties in more than one direction (at the same time), which straightforwardly contradicts quantum-mechanical principles. I discuss this alleged spin-ambiguity, and conclude that (luckily) it does not survive closer scrutiny.

A fundamental objective of philosophy of science is to determine how scientific theories relate to reality. While much attention has been given to the ontological status of concepts like the wave function in quantum mechanics, the status of fiber bundles has received comparatively little discussion. This is unfortunate, given that our most fundamental physical theories can be written in the geometrical language of fiber bundles. More precisely, our currently most successful and fundamental physical theories—namely, the Standard Model of particle physics and general relativity—are gauge theories, and their most general mathematical framework is expressed in terms of fiber bundles. This naturally raises the question of whether fiber bundles are physically real. Although this question is not often explicitly addressed, it has been argued that similarly to how Einstein’s notorious hole argument rules out spacetime substantivalism, a generalized version of the argument rules out fiber bundle substantivalism. In my talk, I will show that by employing the recently established dressing field method, which eliminates gauge redundancy, we can construct a dressed principal bundle that is immune to the hole argument under certain conditions. This sheds new light on fiber bundle substantivalism and deepens our understanding of gauge theories.

The advent of large language models (LLMs) enables machine-assisted semantic analysis of text in an unprecedented manner, bringing into reach new research methods for text-based sciences like philosophy. Despite their immense potential, however, LLMs face crucial limitations, such as the dilution of nuanced philosophical content or missing information about texts that have not been included in the training data. Retrieval-Augmented Generation (RAG) systems address these issues by first setting up an explicit corpus of texts, and then, for each research question, retrieving relevant texts before generating responses on this basis. In this way, RAG systems can crucially enhance precision and depth of the results. This talk explores the potential applications of RAG systems in philosophy, their methodological challenges—including semantic search relevance, corpus selection, and granularity—and their technical implementation. Using a RAG system built on the Stanford Encyclopedia of Philosophy, I illustrate both its promise and the complexities of fine-tuning it for reliable philosophical inquiry.