Machine Learning Group
Practical machine learning through foundational theory
About the Group
The Machine Learning Group at the University of Hamburg studies the foundations of machine learning models and algorithms. Our research aims to understand the structural principles that govern learning methods and to use this understanding to develop robust, principled, and practically effective algorithms.
Modern machine learning encompasses a large and rapidly growing landscape of architectures, training strategies, and optimization techniques. We analyze these methods at a mathematical and algorithmic level to determine which components are essential, which are redundant, and under which conditions learning succeeds or fails. This structural perspective informs the design of improved learning and optimization methods grounded in explicit theoretical insight.
Research Areas
Our work concentrates on:
- Modern learning architectures
Structural analysis of deep models, including transformers and diffusion-based methods, with a focus on inductive biases and architectural constraints. - Optimization and training dynamics
Analysis and development of optimization methods for classical and deep learning, emphasizing stability, convergence, and generalization behavior. - Mathematical foundations of learning algorithms
Differentiation, matrix and tensor calculus, and algorithmic structures underlying modern machine learning. - Principled method design
Development of learning and optimization algorithms with theoretical guarantees and strong empirical performance.
Selected Contributions
Representative contributions include:
- Mathematical characterization of structural position bias in transformer models
- Reliability limits of explainability methods
- Structural analysis of capsule networks and their limitations
- Principled optimization frameworks for machine learning
- Foundational methods for matrix and tensor calculus widely used in research and practice
A full list of publications can be found on the Publications page.