Markus Diesmann
Keynote speakers
- Kenji Doya
- Alon Halevy
- Astrid Prinz
- Andrew Schwartz
- Shankar Subramaniam
- Arthur Toga
Workshop speakers
- Bart ter Haar Romeny
- Uri Eden
- Klaus Linkenkaer-Hansen
- Tim Clark
- Alan Ruttenberg
- Jeffrey Grethe
- Arnd Roth
- Wulfram Gerstner
- Peter Hunter
- Markus Diesmann
- Andrey Semin
- Pietro Liò
- Albert Cardona
- Giorgio Ascoli
- Rolf Kötter
Workshop 4, High performance computing and grid infrastructure for neuroinformatics applications
Markus Diesmann
Title: Perspectives and challenges of large-scale neuronal network simulations
RIKEN Brain Science Institute, Wako, Japan
Abstract: Large-scale simulations of neuronal networks have become a critical tool to integrate the experimental and theoretical findings. At present considerable work is being carried out at the level of local networks of phenomenologically described model neurons. However, in order to increase the explanatory power of the models they need to be extended. First, multi-scale models are required to connect the microscopic level to the mesocopic and macroscopic level where most of the functionally and clinically relevant data are obtained. Moreover, in order to understand the mechanisms of plasticity and neuromodulation, a multi-scale model may also need to include the level of systems biology. Second, because brain functions are distributed over several brain areas, the models need to be large enough to close these functional circuits.
The development of these large- and brain-scale simulations presents a number of technological challenges, such as extending the scalability of simulations with biologically realistic connectivity and heterogeneity from several hundred to tens of thousands of processors. However, the major challenges are sociological and cultural. For example, within a software project the increasing complexity of the code requires the adoption of tools and management strategies. Using the example of the NEST simulation software I illustrate critical phases in the development over the last 15 years and the technologies that have been employed to overcome these problems. The need for a perpetual adjustment of techniques will persist and the talk outlines potential future directions.
Taking a broader scope, it currently appears close to impossible to reproduce published simulation results across simulation tools and research groups, and frighteningly little has been published on the suitability of various simulation techniques for the problems at hand. The reasons have been manifold; most particularly lack of skill of the researchers to report on computer science results and ignorance of funding agencies and reviewers about the neuroinformatics problems of computational neuroscience. This has partly changed now. In the past few years it has become easier to publish on simulation technology and due to large scale projects like FACETS a community is emerging which shares ideas and compares results. Furthermore, INCF has established some activities to overcome this problem. The lack of sustained funding remains. I argue that a culture of comparing and benchmarking needs to emerge. We can only progress if we know how good we are and what is wrong. This holds for correctness of results and appropriate measures of efficiency. We need to learn how to build on the work of others to proceed.
Acknowledgements:
Next-Generation Supercomputer Project of MEXT (Japan), EU Grant 15879 (FACETS), BMBF Grant 01GQ0420 to BCCN Freiburg, Helmholtz Alliance on Systems Biology (Germany).
Bio sketch: Markus Diesmann received his Diploma and PhD in Physics from the University of Bochum (Germany) in 1994 and 2002. The PhD work was conducted at the Weizmann Institute of Science, Rehovot (Israel) and continued at the University of Freiburg (Germany). In 1999 Markus joined the Max-Planck-Institute for Dynamics and Self-Organization, Goettingen (Germany) for a staff position. From 2004 to 2006 he was appointed Juniorprofessor for Computational Neurophysics at the University of Freiburg. Since September 2006, Markus is a Unit Leader at the RIKEN Brain Science Institute, Wako (Japan). His interests include the correlation structure of cortical networks and large-scale simulations.