Meet

Instructor

Prof. Pablo Quesada Barriuso

Biography

Pablo Quesada Barriuso is an assistant professor in the Department of Electronics and Computer Engineering at the University of Santiago de Compostela (Spain). He received a MS in Computer Graphics, Video Games and Virtual Reality in 2010 and an awarded cum laude PhD in Information Technology Research in 2015. He is member of the hyperspectral computing group within the department and member of the High-Performance and Disruptive Computing in Remote Sensing (HDCRS) Working Group of the IEEE GRSS ESI Technical Committee.

Pablo has contributed to the state of the art in remote sensing and high-performance computing with several works mainly focused on feature extraction and efficient schemes of segmentation and classification of hyperspectral images which have been published in top-ranked journals and conferences.

His research interests include multi / hyper spectral remote sensing data processing in low-cost devices focused on feature extraction and analysis of data to reduce transfer time, save storage space, and speed-up data analysis for real time processing by programming parallel applications in shared memory architectures and general-purpose computing on GPU.

Lecture content

Stencil Computation applied to Remote Sensing Hyperspectral Preprocessing on Shared Memory Systems using OpenMP

In parallel programming we frequently find computing patterns that are common in parallel algorithms. An example is the stencil computation where each output element is computed with data from its neighborhood defined by a mask. For this computation the number of operations per pixel is very large. We will see in this use case how to speed up two common pre-processing steps in hyperspectral analysis using OpenMP, from spectral feature reduction using wavelets to morphological profiles by means of the same parallel pattern.”  

Meet

Instructor

Dr. Álvaro Ordóñez Iglesias

Biography

Álvaro Ordóñez received the Ph.D. in Computer Science in 2021, Master's in Big Data Technologies in 2016, and Bachelor's degree in Computer Science in 2015, all from the Universidade de Santiago de Compostela (USC). Currently, he is postdoctoral researcher at CiTIUS (Centro Singular de Investigación en Tecnoloxías Intelixentes) working on remote sensing image processing. He has collaborated in the organisation of the International European Conference on Parallel and Distributed Computing (Euro-Par) and is a member of the High-Performance and Disruptive Computing in Remote Sensing (HDCRS) Working Group.

His research interests are in pre-processing and registration of multi and hyperspectral images in high performance computing, including NVIDIA GPUs and shared and distributed-memory systems.

Lecture content

Comparing different HPC Solutions for Efficient Registration of Multispectral Remote-Sensing Images

Image registration is an essential task in many applications of multispectral remote sensing images. Before any processing, the images must be aligned. This process requires a high computational cost that can be alleviated by carrying out implementations for specialized hardware such as GPUs or shared or distributed memory systems.

Meet

Instructors

Marcel Aach

Biography

Marcel Aach is a Ph.D. student at the Juelich Supercomputing Centre and the University of Iceland. He obtained his M.Sc. in Economathematics from the University of Cologne in 2021. His research interest include large scale machine learning on high-performance computing (HPC) systems with a focus on distributed hyperparameter optimization and neural architecture search.

Rocco Sedona

Biography

Rocco Sedona received the B.Sc. and M.Sc. degrees in information and communications engineering from the University of Trento in 2016 and 2019, respectively. He is member of the ‘‘High Productivity Data Processing’’ (HPDP) research group at the Jülich Supercomputing Centre, Germany. He is currently pursuing the Ph.D. degree in computational engineering at the University of Iceland.

His research interest is mainly in machine learning methods for remote sensing applications, with a particular focus on distributing deep learning models on multiple GPUs of High Performance Computing (HPC) systems.

Lecture content

HPC for Distributed Deep Learning and Hyperparameter Tuning

In this session, we will cover deep learning and how to achieve scaling to high performance computing systems.  We will start the lecture with a presentation of high performance computing system architectures and the design paradigms for HPC software. Furthermore, we give a recap of important machine learning concepts and algorithms. Afterwards, we introduce how deep learning algorithms can be parallelized for supercomputer usage with Horovod. Furthermore, we discuss best practicies and pitfalls in adopting deep learning algorithms on supercomputers and show a practical use case from remote sensing where HPC is used to accelerate training and hyperparameter tuning.