ATML Applied Machine Learning
The JSC Lab Applied Machine Learning applies recent progress in the field of Machine Learning and Artificial Intelligence to topics relevant in science and industry and tailors new approaches to the specific requirements. Depending on the question, our original contributions range from raising fundamental questions, over highly-performing implementations, to methods-centric research. Almost no real-world application of ML is just "data in - train model - deploy model". This is especially true for scientific applications, where data can be scarce, labeling might require costly experiments or it might be just impossible to obtain realistic training data because the experiment still needs to be built. On the other hand, symmetries, physical knowledge, approximate models and other helpful prior knowledge can be available. This mix creates the exciting substrate of our research.
We tackle these challenges with a wide variety of approaches. Deep Learning, statistical learning, self-supervision, differentiable simulation, generative models and equivariant networks are a few examples of techniques we are interested in. Our mothership, the Jülich Supercomputing Centre hosts the largest Supercomputers in Germany. We always ask: When can more computational power improve our methods even further? We want our approaches to benefit from scaling compute to the max. Have you ever tried to run an ML training on 3072 GPUs? We have.
We are at the interface between researchers from various fields and machine learning. Therefore, we are very interested in combining ML methods with domain-specific approaches and/or simulations. We firmly believe that using elements of classical theories such as symmetries can strongly boost the quality of ML methods. Progress in the ML field can also greatly improve classical approaches, for example when using differentiable simulations.
Our lab hosts Helmholtz AI consultant team @ FZJ. The team represents the research field Information and is associated with exciting research domains such as Neuroscience, Quantum computing and Material Science. We support scientists with our AI competence in all phases of their projects: from sketch, through implementation to communication.
As part of the Gauss Centre for Supercomputing, we offer High-Level-Support to potential, novel and actual users of the GCS supercomputers. Again, our support extends from concept to implementation and dissemination of our projects.
The team support the JSC infrastructure in several ways. We maintain the installations of all relevant ML software stacks and make sure that the usage of the supercomputers for ML is smooth. Furthermore, we support HAICORE compute resources where we make sure that Helmholtz association researchers can use JSC's resources for AI-related activities. Our work can be considered part of the support infrastructure as well.
Within Helmholtz AI, researchers can easily apply for our collaboration and support using the Voucher submission system. We highly recommend discussing with us first. Please contact a team member or our central mail address.
Dr. Stefan Kesselheim
Building 16.4 / Room 321