HeAT - A New Versatile Distributed-Parallel Machine Learning Toolkit

Abstract

In this talk we will present the Helmholtz Analytics Toolkit (HeAT), a scientific big data analytics library for HPC systems. Big data analytics, and machine learning in particular, has gained a lot of traction recently with applications ranging from neuroscience to aeronautics. At the heart of many of these methods lies the idea to model even discrete decision processes with continuous functions, whose programmatic handling of course necessitates numerical linear algebra, a prime subject of HPC. However, many classical HPC algorithms are designed with sparsity in mind, a property most often inherited from the locality of a physical description. Since such a locality is missing a priori in most big data scenarios, it poses new challenges on the efficient algorithm design. We will discuss our current efforts to make such algorithms available to the broader public in the HeAT framework.