Hybrid ML (1/2)

by Christian Bauckhage (Fraunhofer IAIS), Lorenz Sparrenberg (University of Bonn)

Friday 7 Nov 2025, 14:15 → 15:45 Europe/Berlin

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

Hybrid Learning (motivated through a simple example) by Christian Bauckhage

Mathematical equations are _the_ modeling tool for the natural sciences. Alas, many if not most problems (e.g. in physics) lead to equations for which there is no closed form solutions. Traditionally, these problems have been addressed using numerical computing but, nowadays, machine learning offers another approach. Of particular interest in this regard are hybrid machine learning models which combine knowledge- and data-driven techniques and we will look at a “simple yet difficult” setting to elaborate on what this means.

Model Inference on Commodity Hardware: The Impact of Quantization on LLMs by Lorenz Sparrenberg

Modern language models comprise many billions of parameters, which makes them difficult to use locally. The solution: quantization, i.e., representing the model weights with lower precision. But how does this work, and how do we measure the impact on our model performance?