Competence Center for Machine Learning Rhine-Ruhr

Machine learning is the basis of the digital transformation. Hence, internationally outstanding research and effective transfer into applications is of the utmost importance for a society. Germany and France aim at a collaboration in machine learning research. In this context, the Competence Centre for Machine Learning Rhine-Ruhr (ML2R), funded by the Federal Ministry of Education and Research (BMBF), is now being launched in Dortmund and Bonn/Sankt Augustin.

The TU Dortmund University, the University of Bonn and the Fraunhofer Institutes for Intelligent Analysis and Information Systems IAIS in Sankt Augustin and for Material Flow and Logistics IML in Dortmund will jointly investigate models of machine learning for their understandability, fairness, traceability and robustness and develop new model classes. Distributed learning, learning from data streams, embeddings, knowledge graphs, modern hardware for machine learning, Bayes' approaches to deep learning and interactive data exploration are some of the topics, according to the motto "a good theory is the best practice".

Dortmund has a history of excellent research in machine learning at the Chair of Artificial Intelligence, Prof. Dr. Katharina Morik. The close connection between theory and practice is a leitmotif that was made clear by the first efficient implementation of the support vector machine, SVM_light, by her doctoral student Thorsten Joachims. Similarly, the successful spin-off RapidMiner illustrates that basic research also facilitates practical developments and applications. The Collaborative Research Center 876, "Machine Learning with Resource Restriction," has been uniting for nearly 8 years machine learning and embedded systems. More than 50 PhD students and postdocs as well as the intensive collaboration with the chair for astroparticle physics and the faculty of statistics provide an agile research environment.

We are hiring

We offer doctoral positions in all key aspects, provided that you have outstanding academic performance, a degree in computer science and previous knowledge of machine learning, e.g. by attending relevant courses during your studies or by publications at the respective conferences.

Please send digital applications with the usual documents stating ML2R in the subject to: office@ls8.cs.tu-dortmund.de

Key Activity: Machine Learning and Resource Restrictions

With the availability of more and more data at different locations through the cost-effective use of embedded systems and sensors, new challenges are posed to machine learning.

The competence center investigates models of machine learning for distributed, severely limited computer architectures and evaluates modern hardware with regard to its suitability for machine learning. In this context, learning methods are examined and, if necessary, adapted with regard to their feasibility on so-called ultra-low-power (ULP) architectures.

If existing devices do not have the necessary functions with low resource requirements, new hardware for machine learning is designed using Field Programmable Gate Arrays (FPGAs), in order to adapt model and hardware to each other. Models are adapted by model compression, regularization, parameter sharing and numerical approximation.

In recent years, Quantum Computing has experienced considerable growth not only in the construction of quantum hardware, but also in the development of quantum algorithms. With the increasing availability of more and more qubits and high fidelity quantum gates, the development of algorithms for understanding the performance of these machines has become more and more important, and the research topics in the competence center are linked to current work on the implementation of probabilistic models with quantum annealing as well as to approximate inference in probabilistic models using quantum annealing.

The Chair of Artificial Intelligence has access to NASA Ames Research Center's D-Wave Quantum Antennas for practical testing of new methods.

Key Activity: Interactive Modeling

Exploration of large amounts of data requires real-time analyses to give the data scientist an overview of the data and help her to form, refine or annotate hypotheses; projections of high-dimensional data, non-negative matrix factorization, spatio-temporal models and outlier detection are unsupervised methods; active Learning involves the user to annotate some data; transfer learning can be used to make suggestions to users for further analysis steps; sampling and counterfactual modeling facilitate interactive tests that then take the user to further analysis steps.

Key Activity: Curating and Certification

The theory of machine learning must show more than runtime and memory requirements; proven guarantees of energy consumption, runtime complexity, memory requirements, error limits, privacy, example complexity of models of learning are important for the deployment of learned models. The insights into resource consumption and proven guarantees can then be converted into comprehensible labels analogous to the energy labels for household appliances. Properties of models and algorithms of machine learning should thus become comprehensible.

Based on the curated data, the components of learning process, i.e., data preprocessing, feature extraction, modeling, model validation and application are to be described in such a way that analysis processes can be easily adapted and partly automatically obtained from existing processes.