Navigation and service

Profile of IAS-8

The division Data Analytics and Machine Learning (IAS-8) researches and develops methods and algorithms for Machine Learning (ML), Data Analytics, Image Processing and Computer Vision. The application focus is on data and problems relevant to Jülich foci, mostly involving tasks from imaging or other gridded data. The goal is to make advances in the methodology of Data Science and to apply these directly with Jülich institutes in the natural and material sciences, to make them usable, and thus directly build a translational bridge. New and further development of methodological approaches for the analysis and evaluation of big data on scalable HPC systems is particularly important here.

The Computer Vision group makes use of machine learning methods, develops them further and brings them into application. Methodologically, however, it addresses more than ML for gridded data. Computer vision, visual-data analytics, and image processing also include data modeling and the associated inference of model parameters, use simulation methods to generate data, and can be applied to even the smallest data sets if domain knowledge is available for data modeling. Thus, they are not only "Big Data" but "All Data", not only "Learning" but also "Inference", not only "HPC", but can also target even the smallest computing resources, currently often smartphones but mostly desktop PCs or workstations. However, even if solutions are to be found for the smallest computers, scalable learning and simulation methods on supercomputers are usually necessary for their development.

The core activities of IAS-8 are thus in the following areas:

  • Further and new development of machine learning methodologies, data analytics, image processing and computer vision,

    • by means of unsupervised and supervised approaches, in particular deep learning by means of artificial neural networks, outlier detection, clustering methods
    • for transparent validation and verification of complex models
    • Incremental approaches for efficient and effective adaptation of models to dynamically changing data and requirements, transfer learning and domain adaptation, as well as to manual adaptations and representation of domain knowledge.
  • Translational research and collaboration with materials and natural science disciplines at the research center,

    • in particular, combining and complementing statistical models with descriptive models.
  • Scalability of algorithms through new algorithmic strategies, filtering techniques, use and combination of multicore CPUs, graphics cards, and distributed systems.