Select your language

HomeAkcePřednáškyLevin Hornischer: Explaining Neural Networks with Reasons

Monday 9. 11. 2026 14:00

lecture | Institute of Philosophy, Czech Academy of Sciences, Jilská 1, Prague (seminar room 124a)

Levin Hornischer: Explaining Neural Networks with Reasons

Organized by the Department of Logic

Detailed information

Invitation pdf

Levin Hornischer (LMU Munich): Explaining Neural Networks with Reasons

Abstract

We propose a new interpretability method for neural networks, based on a novel mathematico-philosophical theory of reasons. Our method computes a so-called reasons vector for each neuron, via the neuron's activation across a range of inputs. With our theory, we next compute how strongly this reasons vector speaks for human-understandable propositions: e.g., 'depicting digit 2' or 'having negative sentiment'. The theory also dictates how to compose interpretations of individual neurons to interpret groups thereof. Our method combines logical and Bayesian perspectives and accounts for polysemanticity (i.e., that a single neuron can figure in multiple concepts). We show both theoretically and experimentally that our method: (1) applies uniformly across architectures and modalities, (2) is scalable, (3) computes strengths for any proposition and not just for output features, (4) is faithful under interventions, and (5) improves robustness and fairness. 

This is joint work with Hannes Leitgeb, Cordelia Berz, and Hosea von Hauff. A preprint is available here: https://arxiv.org/abs/2505.14424.