Research

The Zuse School ELIZA advances scientific excellence in machine learning-driven AI through collaboration between academic and industrial fellows. ELIZA addresses critical AI challenges—robustness, adaptability, and explainability—by connecting foundational research with practical applications via ELLIS Units and industry partnerships. Our doctoral program features co-supervision, with PhD students jointly mentored by fellows across ELIZA sites or from academic and industrial backgrounds. Master's scholars receive dedicated supervision from ELIZA fellows, supporting their early research endeavors. By attracting premier national and international AI talent, ELIZA strengthens Germany's position as a leading destination for researchers and innovators.

Foundations of Machine Learning

The "Foundations of Machine Learning" research area at ELIZA delves into the theoretical underpinnings and mathematical principles that drive the development of machine learning algorithms and systems. This area crucially advances the field of Artificial Intelligence by establishing a robust scientific basis for how machines learn from data. It encompasses a wide range of topics aimed at understanding and improving the fundamental capabilities of learning algorithms. Notably, this area includes ML-driven disciplines such as computer vision, natural language processing and robot learning, thereby broadening the theoretical focus to more applied domains within AI. PhD students associated with ELIZA are actively engaged in this research area, contributing to a deeper comprehension of learning processes and the development of novel theoretical frameworks.
Applications in Autonomous Systems

Autonomous systems represent one of the most transformative areas of AI. ELIZA fellows working in this research area collaborate with industry leaders to enhance robotics, autonomous vehicles, and adaptive control systems, addressing key challenges such as reliability and real-world applicability. The "Applications in Autonomous Systems" research area at ELIZA focuses on the development and application of machine learning techniques to enable intelligent and independent behaviour in various systems. Work in this area leverages advancements in machine learning to create autonomous agents capable of perceiving their environment, making decisions and acting effectively without constant human intervention. PhD students within ELIZA are actively involved in a diverse range of projects that aim to advance the capabilities of autonomous systems in various real-world scenarios.
Machine Learning Systems

The Machine Learning Systems research area at ELIZA focuses on the engineering and design of efficient, scalable and robust systems for running and deploying machine learning models. This area bridges the gap between theoretical machine learning algorithms and their practical implementation in real-world applications. This research is essential for enabling the widespread adoption of AI technologies across diverse sectors and ensuring that machine learning innovations can deliver meaningful impact in production environments.
Trans-disciplinary Applications

Machine learning holds immense potential for scientific discovery across disciplines. ELIZA scholars apply AI techniques to fields such as life sciences and physics, enabling breakthroughs in medical imaging, computational biology, and complex simulations. The "Transdisciplinary Applications" research area at ELIZA applies machine learning methodologies and techniques to address challenges and foster innovation across a diverse range of scientific disciplines. This area aims to exploit new technological-scientific areas or usage options of AI, thus going beyond classical subject or application boundaries.

Significant Runtime Reduction for Asynchronous Multi-Fidelity Optimization on Zero-Cost Benchmarks

by Shuhei Watanabe, Neeratyoy Mallik, Edward Bergman, Frank Hutter

DOI:

This paper introduces a Python package that enables efficient parallel benchmarking of asynchronous multi-fidelity hyperparameter optimization (HPO) on zero-cost benchmarks. Traditional zero-cost benchmarks remove the need for actual model training, but in parallel/asynchronous settings, each worker still has to wait for simulated runtimes to preserve correct evaluation order—leading to major slowdowns. The proposed solution calculates the correct return order using file system synchronization, eliminating unnecessary waiting and achieving over 1000× speedup compared to naïve approaches.
Python Wrapper for Simulating Multi-Fidelity Optimization on HPO Benchmarks without Any Wait

by Shuhei Watanabe

DOI: https://doi.org/10.48550/arXiv.2305.17595

This paper introduces a Python wrapper that eliminates waiting time in multi-fidelity optimization using tabular or surrogate benchmarks. The wrapper solves the fundamental problem where simulators of asynchronous HPO must wait for actual training runtime at each iteration to maintain correct evaluation order. The solution forces each worker to wait only 10^-2 seconds instead of several hours while yielding exactly the same evaluation order as real experiments. This enables efficient parallel HPO with zero-cost benchmarks, achieving over 1000× speedup compared to traditional approaches.
Python Tool for Visualizing Variability of Pareto Fronts over Multiple Runs

by Shuhei Watanabe

DOI: https://doi.org/10.48550/arXiv.2305.08852

This paper introduces a Python package for visualizing the variability of Pareto fronts over multiple runs using empirical attainment surfaces (EAS). The tool addresses a key limitation in multi-objective optimization evaluation: while Pareto fronts are often shown from single runs without considering variability from random seeds, this makes performance stability evaluation difficult. The package provides uncertainty visualization for Pareto fronts, enabling researchers to assess the robustness and reliability of multi-objective optimization methods across multiple independent runs.

Research

Foundations of Machine Learning

Applications in Autonomous Systems

Machine Learning Systems

Trans-disciplinary Applications

Recently published

Significant Runtime Reduction for Asynchronous Multi-Fidelity Optimization on Zero-Cost Benchmarks

Python Wrapper for Simulating Multi-Fidelity Optimization on HPO Benchmarks without Any Wait

Python Tool for Visualizing Variability of Pareto Fronts over Multiple Runs