% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@PHDTHESIS{Aach:1043684,
author = {Aach, Marcel},
title = {{P}arallel and {S}calable {H}yperparameter {O}ptimization
for {D}istributed {D}eep {L}earning {M}ethods on
{H}igh-{P}erformance {C}omputing {S}ystems},
school = {University of Iceland},
type = {Dissertation},
reportid = {FZJ-2025-02982},
isbn = {978-9935-9807-8-6},
pages = {172p},
year = {2025},
note = {Additional Grant: Verbundprojekt: NXTAIM - NXT GEN
(01.01.2024-31.12.2026); Dissertation, University of
Iceland, 2025},
abstract = {The design of Deep Learning (DL) models is a complex task,
involving decisions on the general architecture of the model
(e.g., the number of layers of the Neural Network (NN)) and
on the optimization algorithms (e.g., the learning rate).
These so-called hyperparameters significantly influence the
performance (e.g., accuracy or error rates) of the final DL
model and are, therefore, of great importance. However,
optimizing these hyperparameters is a computationally
intensive process due to the necessity of evaluating many
combinations to identify the best-performing ones. Often,
the optimization is manually performed. This Ph.D. thesis
leverages the power of High-Performance Computing (HPC)
systems to perform automatic and efficient Hyperparameter
Optimization (HPO) for DL models that are trained on large
quantities of scientific data. On modern HPO systems,
equipped with a high number of Graphics Processing Units
(GPUs), it becomes possible to not only evaluate multiple
models with different hyperparameter combinations in
parallel but also to distribute the training of the models
themselves to multiple GPUs. State-of-the-art HPO methods,
based on the concepts of early stopping, have demonstrated
significant reductions in the runtime of the HPO process.
Their performance at scale, particularly in the context of
HPC environments and when applied to large scientific
datasets, has remained unexplored. This thesis thus
researches parallel and scalable HPO methods that leverage
new inherent capabilities of HPC systems and innovative
workflows incorporating novel computing paradigms. The
developed HPO methods are validated on different scientific
datasets ranging from the Computational Fluid Dynamics (CFD)
to Remote Sensing (RS) domain, spanning multiple hundred
Gigabytes (GBs) to several Terabytes (TBs) in size.},
cin = {JSC},
cid = {I:(DE-Juel1)JSC-20090406},
pnm = {5111 - Domain-Specific Simulation $\&$ Data Life Cycle Labs
(SDLs) and Research Groups (POF4-511) / RAISE - Research on
AI- and Simulation-Based Engineering at Exascale (951733) /
nxtAIM - nxtAIM – NXT GEN AI Methods (19A23014l)},
pid = {G:(DE-HGF)POF4-5111 / G:(EU-Grant)951733 /
G:(BMWK)19A23014l},
typ = {PUB:(DE-HGF)3 / PUB:(DE-HGF)11},
doi = {10.34734/FZJ-2025-02982},
url = {https://juser.fz-juelich.de/record/1043684},
}