This title appears in the Scientific Report :
2023
Leveraging L2L for intelligent parameter space exploration on HPC using GPUs for TVB
Leveraging L2L for intelligent parameter space exploration on HPC using GPUs for TVB
The Virtual Brain (TVB; Sanz Leon et al. 2014) is neuroinformatics platform for full brain network simulations using biologically realistic connectivity. In order to abstract implementation complexity of TVB’s whole brain network models, the model generator RateML (Vlag et al. 2022) can be utilized....
Saved in:
Personal Name(s): | van der Vlag, Michiel (Corresponding author) |
---|---|
Contributing Institute: |
Jülich Supercomputing Center; JSC |
Imprint: |
2023
|
Conference: | Concluding Event of the Human Brain Project, Jülich (Germany), 2023-09-12 - 2023-09-13 |
Document Type: |
Conference Presentation |
Research Program: |
SimLab Neuroscience Human Brain Project Specific Grant Agreement 3 Domain-Specific Simulation & Data Life Cycle Labs (SDLs) and Research Groups |
Publikationsportal JuSER |
The Virtual Brain (TVB; Sanz Leon et al. 2014) is neuroinformatics platform for full brain network simulations using biologically realistic connectivity. In order to abstract implementation complexity of TVB’s whole brain network models, the model generator RateML (Vlag et al. 2022) can be utilized. One of its outputs is a model and simulator object for the Compute Unified Device Architecture (CUDA) platform, enabling parameter exploration by making use of the parallel architecture of the GPU. In order to find certain brain regimes for this model intelligently, a scalable hyper parameter optimization framework called L2L, implementing the concept of learning to learn, can be instantiated to do parameter space exploration in an automated and parallel fashion (Yegenoglu et al. 2022). L2L is agnostic to the process to be optimized, called optimizee, its optimizers use population based decision algorithms, and simulations can run embarrassingly parallel. A simulation which has been used in a study for scale-integrated understanding of conscious and unconscious brain states and their mechanisms (Goldman et al. 2021), is used as a demonstrator to efficiently find desired compound parametrizations, using the structure to function correlation as fitness. This workflow (which will be presented), from model generation to intelligent parameter exploration, which is an integral part of WP1 showcase 1, makes the computational infrastructure on the supercomputer accessible to scientists. |