Hybrid ML (2/2)

Friday 14 Nov 2025, 14:15 → 15:45 Europe/Berlin

DO: JvF25/3-303 | BN: b-it/1.047

Unpacking Permutations by Zorah Lähner

The talk will give an overview over the research done in the Geometry in Machine Learning group and in detail one method to represent permutation matrices in neural networks. It is a way to convert the binary, high dimensional and highly constrained permutation matrix, which in general cannot be optimized properly in neural networks, in two real-valued matrices of much lower dimensionality. The construction can be derived from purely geometric arguments using the sphere packing problem which we will do in the lecture.

Characterizing Functional Connectivity Subnetwork Contributions in Narrative Classification with Shapley Values by Aurora Rossi

This talk explores brain activity during narrative tasks using functional Magnetic Resonance Imaging (fMRI). The evolving functional connectivity of the brain is modeled as a temporal network, and machine learning methods are applied to classify different narrative aspects. Beyond classification, Shapley values are used to quantify the contribution of specific brain subnetworks to narrative comprehension, providing novel insights into the brain’s functional organization.