Levin Hornischer
LightNews LightNews
banner
levinhornischer.bsky.social
Levin Hornischer
@levinhornischer.bsky.social
210 followers 130 following 7 posts
Assistant professor at LMU Munich, MCMP (Munich Center for Mathematical Philosophy). He/him.
Working on: foundations of AI, logic, dynamical systems, semantics, epistemology.
https://levinhornischer.github.io/
Posts Replies Media Videos
levinhornischer.bsky.social
New paper in JAIR, with Zoi Terzopoulou: 'Learning How to Vote with Principles: Axiomatic Insights Into the Collective Decisions of Neural Networks'

❓Can neural nets find new voting rules to aggregate preferences?
💡Yes, by optimizing for axioms!

jair.org/index.php/ja...
philpapers.org/rec/HORLHT
The preferences concerning three alternatives (apple, banana, cherries) of three stylized individuals are fed into a stylized neural network. The network outputs a ranking of the three alternatives as the aggregated preference of the group. It selects the top ranked alternative, in this case the apple. In the bottom right, there are two stylized scientists wondering if this network is 'anonymous', 'biased', and 'cycle-free' in its preference aggregation.
August 6, 2025 at 5:55 PM
levinhornischer.bsky.social
New preprint, with Hannes Leitgeb @lmu-mcmp.bsky.social: "Explaining Neural Networks with Reasons".

➡️We propose a new faithful and scalable interpretability method for neural networks.
💡Based on a novel mathematico-philosophical theory of reasons.

arxiv.org/abs/2505.14424
philpapers.org/rec/HORENN
Thousands of small dots in 10 different colors on a white background. Dots with the same color form clusters. A dot represents an input-label pair ('possible world'). Close-by dots are possible worlds that are similar according to the neural network's reasons structure ('internally similar'). The fact that they form monochromatic clusters means that internally similar worlds typically are also externally similar, i.e., have the same label. In this case, there are 10 labels represented by the 10 colors. So the neural network's reasons structure matches that of the world.
May 21, 2025 at 7:44 PM
levinhornischer.bsky.social
New paper in Notre Dame J. Formal Logic: 'Iterating Both and Neither: With Applications to the Paradoxes'

❓What if we keep on adding new truth-values 'neither a nor b' and 'both a and b'?
➡️Fun math and fresh ideas for paradoxes!

Paper: doi.org/10.1215/0029...
Preprint: philpapers.org/rec/HORIBA
The algebra of truth values obtained after iterating 'both' and 'neither' two times. It consists of 16 truth values shown as black dots on a white background. Some are connected by solid lines, which indicates that, in the algebra, one is less-or-equal to the other. There is a small text next to each node describing the truth-value. For example, the node closest to the bottom left corner says { {0,1}_n, {0} }_n. This is the following truth value: Neither 'neither true nor false' nor 'false'. The other nodes are variations thereof.
March 3, 2025 at 1:14 PM