Institute for Advanced Simulation (IAS)
Servicemeu
- Deutsch
- English
-
Search
Institute for Advanced Simulation (IAS)
The team "Multiscale Simulation and Architectures'" in the SimLab Neuroscience is currently involved with computational neuroscience projects for the simulation and analysis of spiking neuronal networks at several space-and-time scales, from the morphologically detailed to coarse-grain full-brain simulation involving both short-term plasticity and long-term evolution of connectivity, in addition to the development of methods for high-throughput optimization and search in this context. Two computing time allocations are closely associated with these projects: ``Parameter fitting for The Virtual Brain using Bayesian Inference'' (TVB, CJJSC35, 2.5M Ch), and ``Adaptive parameterization of structural plasticity models in neural network simulations'' (SP, CJJSC34, 1.2M Ch).
In the TVB project, we have been contributing to the development of an HPC version of the ``The Virtual Brain'' code and the application of Bayesian inference techniques to derive model parameters from large sets of empirical results.
The large ensembles involved are embarrassingly parallel, but the kernels themselves have been ported to take advantage of CUDA and many-core vectorization, while applying hyperparameter inference techniques to the ensembles. This project has also contributed to the Human Brain Project and The VirtualBrainCloud H2020 grants with core infrastructural components.
In the SP project, we have been extending work previously done on the slow evolution of structural plasticity for NEST simulations, accessing network behavior at new time scales, and contributing to testing of increasingly scalable implementations for HPC.
Learning-2-Learn techniques are used to optimize plasticity hyperparameters at scale using the JuPeX HPC framework developed in this context as a driver in the development of high-throughput optimization techniques. Specifically, we implement hyperparameter multiobjective optimization of the homeostatic rules of the structural plasticity algorithm for different sizes of spiking networks using fitness measures of firing rate, shape of the power spectrum and synchronization.