Session

DescriptionDeep learning (DL) has the capability to transform many scientific problems, including COVID-19 research, cancer studies, and a wide range of energy sciences. Ever more complex couplings of data sources, analysis, and learning methods demand additional resources in computing capability, data movement, and storage. In practice, the growing scientific usage of DL faces technical and engineering challenges in the exascale era. This minisymposium will illustrate two different learning-based applications and two related systems software contributions to DL at exascale. Exascale challenges for DL include scaling models, scaling workflows, integrating methods, integrating software, and down-scaling results for return to users with varying levels of available computing power. The presenters of this minisymposium are developing a suite of software to support scalable DL on exascale supercomputing resources. The software suite includes capabilities for rapid prototyping and application development, portability across machine types and scales, support for typical DL workflow patterns such as error analysis and hyperparameter optimization, and hierarchical parallelism via distribution of larger models across multiple nodes. We will also present a range of research results in the usage of deep learning ensembles, including recursive search patterns used to look for outliers in experimental data.