Helmholtz AI at Forschungszentrum Juelich – Implementation of AI Methods on High Performance Computing Systems

Abstract

Helmholtz AI at FZJ addresses the current transformation of science regarding different aspects of digitization especially at the overlap of AI with high-performance computing (HPC), and neuroscience. The unit is built on the long and intense interdisciplinary partnership between Juelich Supercomputing Centre (JSC) and Institute of Neuroscience and Medicine, Structural and functional organisation of the brain (INM-1) within the Helmholtz Programmes ‘Decoding the Human Brain’ and ‘Supercomputing and Big Data’, which has already become the main driver of the European research flagship “Human Brain Project”. The focus of Juelich’s Helmholtz AI unit will be on robust deep learning methods for high-resolution scientific image analysis, as well as on large-scale, self-organized continual learning transferable across different tasks and domains. Driven by high-throughput data acquisition and the ambition to quickly transfer learned knowledge across tasks and scientific domains, the implementation of AI methods on large-scale high performance computing systems is a key aspect of this work.
Helmholtz AI at FZJ will be implemented with a research group at INM-1 and a tandem of groups at JSC that cover both research and research support. The research group at INM-1 will focus on Biomedical Computer Vision, especially deep learning methods for analyzing large and complex scientific image data with limited availability of training examples. Driven by continuously increasing image resolutions and data volumes in high-throughput settings, methods for distributed operation on HPC systems will be developed. To ensure a close exchange of research and support, the team at JSC consists of the research-oriented Cross Sectional Team Deep Learning (CST-DL) and the software development and research support-oriented High Level Support Team (HLST). The research focus and long-term agenda driven by JSC in the field of AI will be on enabling large-scale self-organized continual learning in multi-task scenarios. This line of research will create methods capable of growing generic models from incoming streams of data, extracting knowledge and skills quickly transferable across different tasks and domains - a still grand, open scientific question. The activity will be strongly dedicated to open science and open source software, making all the results transparent and all the tools available to scientific communities and public.

Full text

Our Partner


The Team @FZJ
Morris Riedel
Head of Research Group High Productivity Data
Processing, Co-Head of CST-DL
Jülich Supercomputing Centre (JSC)
Associated Professor at University of Iceland
m.riedel@fz-juelich.de
Timo Dickscheid
Head of Big Data Analytics Group
Institute of Neuroscience and Medicine,
Structural and functional organisation of the brain (INM-1)
t.dickscheid@fz-juelich.de
Jenia Jitsev
Co-Head of CST-DL
Institute for Advanced Simulation (IAS)
Jülich Supercomputing Centre (JSC)
j.jitsev@fz-juelich.de
Susanne Wenzel
Big Data Analytics Group
Institute of Neuroscience and Medicine,
Structural and functional organisation of the brain (INM-1)
s.wenzel@fz-juelich.de
Research Group
T. Dickscheid
Coordination S. Wenzel
Large Scale AI
M. Riedel, J. Jitsev Representation
Communication
Junior Group
NN
(Research)
Junior Group
NN
(Research)
CST Deep Learning
J. Jitsev/M.Riedel
(Research)
High Level support Team
NN
(Support)
Joint Artificial Intelligence Machine Learning Lab (JOAIML)
International Collaborations, CIFAR, MILA, McGill, Iceland, NCSA, Israel
Institute for Neuroscience and
Medicine INM-1
Jülich Supercomputing Centre JSC
Head of Helmholtz AI Local
M. Riedel
Local
Research field Information
Organisational Structure
Deep Networks for
Hyperspectral Image
Analysis
Haut, M. & Cavallaro, C. &
Riedel, M., IEEE Transactions
on Geoscience & Remote
Sensing 2019
JUWELS – Jülich
Wizard for
European
Leadership
Science Modular
Supercomputer
Jülich Supercomputing Centre: Large-Scale Continual Learning
transferable across multiple domains and tasks
High Level Support Team (HLST) and Cross-Sectional Team Deep Learning (CST-DL)
Large-scale, self-organizing continual learning for multi-task scenarios
Simulation-learning closed-loops for physics-aware deep learning
Enabling hyperspectral satellite image analysis via distributed deep learning to understand climate change
Cooperation with scientific domains: earth sciences, nanoscale molecular manipulation, plasma physics,
magnetospheric physics, turbulence dynamics, remote sensing.
Explore innovative modular supercomputing approaches (e.g.,with neuromorphic and quantum
computing)
Remote Sensing Big Data Classification
with High Performance Distributed Deep
Learning
R Sedona, G Cavallaro, J Jitsev, A Strube, M Riedel, et al
Remote Sensing 11 (24), 3056
Approching Remote Sensing Image Classification with
Ensembles of Support Vector Machines on the D-Wave
Quantum Annealer; G. Cavallaro, D. Willsch, M. Willsch, K.
Michielsen and M. Riedel, IEEE International Geoscience and
Remote Sensing Symposium, 2020 (submitted)
Steering Board
Jülich Supercomputing Centre
Leveraging innovative modular supercomputing for
distributed training of large heterogeneous AI models
Support Vector Machines on
the D-Wave
Quantum Annealer
Research Group at Institute for Neuroscience and Medicine:
AI methods for building ultrahigh resolution human brain models
Fertilizing brain-inspired AI research. As human brain
models generate insight into layer structure, connectivity,
and temporal dynamics, they can reduce the search space
for novel architectures. The research group interacts with
partners in brain-inspired AI to help finding architectures
with faster learning performance and better
generalizability.
Development of AI methods for microscopic image
analysis and 3D reconstruction enables the
construction of next generation, high-resolution
human brain models.
Institute of Neuroscience & Medicine: Structural and functional
organisation of the brain (INM-1)
INM-1 develops a 3D-model of the human brain which considers cytoarchitecture,
connectivity, molecular structure as well as genetics and function, employing AI and Big Data
methods on High Performance Computers for data analysis.
Human brain mapping in
histological scans using Deep
Learning
Spitzer, Amunts, Harmeling, Dickscheid,
MIDL 2018
Spitzer, Kiwitz, Amunts, Harmeling,
Dickscheid MICCAI 2018
High-resolution 3D reconstruction
from microscopic scans
Dickscheid, Amunts et al.: Towards 3d
reconstruction of neuronal cell distributions
from histological human brain sections.
From Clouds and Big Data to Exascale and
Beyond, 2019
Microstructural object
detection using Deep
Learning
ethods
on
High
Performance
Cooperation