Deep learning has been already successfully applied in many areas of science and industry. Since we are dealing often with extremely large data or very complex neural network architectures, parallelization of deep learning algorithms and frameworks is becoming more and more important. These solutions can no longer be processed on commodity hardware with the high requirement of data security; this is where HPC comes in. When going from classical artificial intelligence (AI) to high-performance AI, we need to ensure that HPC is ready for this endeavour. Thus, today's HPC centers need to provide seamless workflows to enable analytics and deep learning solutions, so that data scientists can fully exploit the performance of HPC systems. In this paper, we demonstrate methodologies for applying deep learning on HPC, and how AI techniques can successfully be integrated with classical simulation codes (e.g. to achieve better accuracy). Furthermore, we present an overview about training neural networks on HPC while successfully leveraging data, model, pipeline and hybrid parallelism. Finally, we adopt these techniques for two use cases: (i) novel hybrid workflow to combine a multi-task neural network with a typical FEM simulation to determine material characteristics, and (ii) segmentation of high-resolution satellite images to identify rice paddies without manual labelling.
%0 Book Section
%1 Zhong2023
%A Zhong, Li
%A Shcherbakov, Oleksandr
%A Hoppe, Dennis
%A Resch, Michael
%A Koller, Bastian
%B Data Science in Applications
%C Cham
%D 2023
%E Dzemyda, Gintautas
%E Bernatavicien\.e, Jolita
%E Kacprzyk, Janusz
%I Springer International Publishing
%K hlrs myown
%P 233--252
%R 10.1007/978-3-031-24453-7_11
%T Towards Seamless Execution of Deep Learning Application on Heterogeneous HPC Systems
%U https://doi.org/10.1007/978-3-031-24453-7_11
%X Deep learning has been already successfully applied in many areas of science and industry. Since we are dealing often with extremely large data or very complex neural network architectures, parallelization of deep learning algorithms and frameworks is becoming more and more important. These solutions can no longer be processed on commodity hardware with the high requirement of data security; this is where HPC comes in. When going from classical artificial intelligence (AI) to high-performance AI, we need to ensure that HPC is ready for this endeavour. Thus, today's HPC centers need to provide seamless workflows to enable analytics and deep learning solutions, so that data scientists can fully exploit the performance of HPC systems. In this paper, we demonstrate methodologies for applying deep learning on HPC, and how AI techniques can successfully be integrated with classical simulation codes (e.g. to achieve better accuracy). Furthermore, we present an overview about training neural networks on HPC while successfully leveraging data, model, pipeline and hybrid parallelism. Finally, we adopt these techniques for two use cases: (i) novel hybrid workflow to combine a multi-task neural network with a typical FEM simulation to determine material characteristics, and (ii) segmentation of high-resolution satellite images to identify rice paddies without manual labelling.
%@ 978-3-031-24453-7
@inbook{Zhong2023,
abstract = {Deep learning has been already successfully applied in many areas of science and industry. Since we are dealing often with extremely large data or very complex neural network architectures, parallelization of deep learning algorithms and frameworks is becoming more and more important. These solutions can no longer be processed on commodity hardware with the high requirement of data security; this is where HPC comes in. When going from classical artificial intelligence (AI) to high-performance AI, we need to ensure that HPC is ready for this endeavour. Thus, today's HPC centers need to provide seamless workflows to enable analytics and deep learning solutions, so that data scientists can fully exploit the performance of HPC systems. In this paper, we demonstrate methodologies for applying deep learning on HPC, and how AI techniques can successfully be integrated with classical simulation codes (e.g. to achieve better accuracy). Furthermore, we present an overview about training neural networks on HPC while successfully leveraging data, model, pipeline and hybrid parallelism. Finally, we adopt these techniques for two use cases: (i) novel hybrid workflow to combine a multi-task neural network with a typical FEM simulation to determine material characteristics, and (ii) segmentation of high-resolution satellite images to identify rice paddies without manual labelling.},
added-at = {2023-04-27T20:11:53.000+0200},
address = {Cham},
author = {Zhong, Li and Shcherbakov, Oleksandr and Hoppe, Dennis and Resch, Michael and Koller, Bastian},
biburl = {https://puma.ub.uni-stuttgart.de/bibtex/25c02e98c73a14c6b7f97dca8f9cf6055/hopped},
booktitle = {Data Science in Applications},
doi = {10.1007/978-3-031-24453-7_11},
editor = {Dzemyda, Gintautas and Bernatavi{\v{c}}ien{\.{e}}, Jolita and Kacprzyk, Janusz},
interhash = {553bdde18bfcbe99d9159af5635e6137},
intrahash = {5c02e98c73a14c6b7f97dca8f9cf6055},
isbn = {978-3-031-24453-7},
keywords = {hlrs myown},
pages = {233--252},
publisher = {Springer International Publishing},
timestamp = {2023-04-27T20:11:53.000+0200},
title = {Towards Seamless Execution of Deep Learning Application on Heterogeneous HPC Systems},
url = {https://doi.org/10.1007/978-3-031-24453-7_11},
year = 2023
}
