Dear Colleagues, please help us diffusing this post-doc offer as much as possible. Thank you Marco Congedo, PhD http://www.novatecheeg.com/Congedo/MC_Home.html Post-Doc Position Where: Paris, PIC, Paris Interdisciplanary Center Centre de Reflexions Interdisciplinaires (CRI) Faculte de Medecine Universite Rene Descartes-Paris V, Site Cochin Port-Royal 24, rue du Faubourg St Jacques 75014 PARIS tel : 01.44.41.22.69 When: As soon as possible Salary: 1800€/month net minimum (more depending on experience) Contact: bernard.hennion@xxxxxxxxxxxxxxxxxx Toward less-blind source separation for brain computer interfaces Supervisors: Bernard HENNION, PhD, France Telecom Research and Development (FTR&D) Ivo SBALZARINI, PhD, Assistant Professor, ETH Zurich, Computational Biophysics Lab 1. Background Today, different blind source separation techniques are used to analyze non-invasive EEG or MEG signals in brain computing interfaces. All of them are based on the hypothesis of exact knowledge of the direct transmission matrix through the gray matter. This matrix describes the propagation of the electromagnetic waves from the active brain regions to the locations of the measurement points. The signals describe the activation of localized electromagnetic dipoles, resulting from the cumulative, synchronous, and simultaneous action of several tens of thousands of parallel oriented neurons, as, e.g., seen in cortical columns. In practice, the direct transmission matrix is, however, unknown, such that the hypothesis underlying blind source separation is hardly true. In addition, a good part of what is usually called "measurement noise" actually contains a significant proportion of signal energy, owing nothing to hazard, but rather containing the neural signals of multiple clusters of small numbers of neurons, or neurons of random directional orientation, such that their signal contributions cannot be accurately detected by the measurement on the skull surface. 2. Research proposal In the present study, we propose to use a powerful direct numerical simulation technique, running on a large number of parallel processors, to simulate the accurate intra-skull propagation of electromagnetic waves. Our simulations will be directly based on first principles from physics (electromagnetics, wave propagation, material properties) and we will use actual and accurate 3D geometries from real brains and skulls as determined by tomographic imaging. This will allow, for the first time, to explicitly account for effects such as smearing on internal faces of skull bones, anisotropy of the brain matter, different physical diffusion parameter values of white matter, grey matter, cerebellum trunk, meninges, cephalo-rachidian liquid, bones, blood circulation, muscle activity signals, skin effects, effects of the convoluted geometric shape of the brain cortex, cross-talk between measurement points, and the exact positioning of the measurement points. In addition, the interface effects at each material frontier that the electromagnetic waves cross can be investigated and accounted for. The proposed physical direct (forward) simulations allow determining an accurate direct transmission matrix, describing the signal propagation from brain voxels to scalp pixels. The simulations are directly based on the proper physical laws of electromagnetic wave propagation through the "details" of intra-skull anatomical geometries. Due to the large number of brain voxels (250 000 with an edge-length of 8mm) and skull pixels (230 000 with an edge-length of 1mm), the complex geometry of the brain, and the different involved materials, the simulations will require the use of parallel supercomputing resources with hundreds of processors. We will thus implement the simulation program on the basis of an existing parallel programming framework, called the PPM Library [1]. This library allows automatic parallelization of scientific simulations and effectively reduces the code development time by several orders of magnitude (as an example, a parallel discrete element simulation for granular flow was implemented within 2 person-weeks, whereas the normal development time for such a code would have been 2 person-years). The PPM library has demonstrated parallel efficiencies and computational speeds that meet or exceed the present state of the art in many fields, including computational fluid dynamics, molecular dynamics, computational cell biology, and astrophysics [1]. The present project will greatly benefit from the use of this library. In order to ensure accurate and efficient simulations, we propose to use adaptive multi-resolution particle methods to discretize the governing equations [2,5]. Particle methods have unique characteristics [5] that allow simulations in complex geometries [3], simulations on multiple spatial scales (i.e. to achieve finer resolution in the cortex than in the white matter) [2,5], and have favorable robustness properties with respect to noise and instabilities. The direct transmission matrix for a specific brain geometry (i.e., person) can then be identified using the proposed simulations. Therefore, we will independently simulate the measured read-out on the skull (at the positions of the measurement points) for point-source signals originating from each of the brain voxels. This yields the individual point transmission matrices that can then be superimposed in order to obtain the full direct transmission matrix (Green’s function superposition principle). This accurate transmission matrix can be inverted numerically in order to obtain its inverse. Multiplication of the EEG or MEG measurements with this inverse then directly yields the activity distribution in the whole volume of the brain. The spatial resolution will initially be 8mm, but can later be improved. The achievable resolution is only limited by the available computer resources. Since the PPM library has been successfully used to perform simulations on up to 1024 processors using up to 1 billion computational elements [1], unprecedented resolutions will be possible. We will validate our simulations using experimental data. These will involve both invasive (brain) and non-invasive (skull) EEG measurement as, e.g., collected from epileptic patients prior to surgery. Application of the determined transmission matrix to the invasive measurements should recover the non-invasive signals. Comparison of this outcome with the actually measured non-invasive signals then enables us to compute differences between the simulation model and reality. Using a black-box optimization strategy such as, e.g., Evolution Algorithms [4], we will then adjust the material constants and simulation parameters in the computer (but not the actual physical model), such that the simulations match the experiment. The so-determined constants and parameters constitute a valuable outcome of the proposed project by themselves, as they are largely unknown for brain/skull material today. In addition, experiments and simulations will use the exact same brain geometry (individualized for a specific person). This will allow, for the first time, the study of inter-personal variations in the physical parameters and the brain signal propagation dynamics. 3. Project situation and significance The present project is highly interdisciplinary and situated within the Openvibe project of real-time 3D analysis and visualisation of brain activity from scalp signals. The long-term goal of the Openvibe project is to ameliorate mental command comfort. More accurate transmission matrices, as provided by the present project, are of crucial importance in achieving this goal. The following scientific fields and expertises are involved in the present project: - High-performance parallel computing/grid computing and software engineering (see point 4 of FTR&D/INSERM agreement) - Development of adaptive multi-resolution numerical methods for the simulation of complex dynamic systems - Precise modelling of living matter, better understanding of brain activity and the process of thought (see point 10 of FTR&D/INSERM agreement) - Brain computing interface techniques (Inverse mathematic problem, sLoreta, stochastic analysis and de-correlation of independent sources) - Mental control comfort (As an example, infrared remotes were, at their origin, designed for handicapped persons. Nowadays, every piece of home-electronic equipment uses them…) - Accessibility ergonomics - France Myopath Association, AFM as a research partner. The present project requires a union of skills and scientific expertise in the following fields: - Ergotherapeut (AFM Claude Dumas, Ornella PLOS PhD) - Cognitive ergonomy (Accessibility, Denis Chêne, FTR&D) - Neuroscience and neural biology (Olivier Bertrand, Jean-Philippe Lachaux, Inserm Lyon, Marco congedo, INPG) - Mathematics: probability and stochastic signals analysis (Christian Jutten, Cedric PhD, INPG) - Physics, electromagnetism, and interface effects (Nelido Gonzalez-Segredo, proposed postdoctoral fellow) - Brain physiology, psychiatry, anatomy, EEG/MEG measurement (Ms. Garnero, Pitié Salpêtrière, physiciens du CRI) - Parallel numerical simulations, (Bernard Hennion FTR&D/CRI, Ivo F. Sbalzarini ETH/CRI, Birte Schrader ETH) - Parallel software engineering and programming (Ivo Sbalzarini ETH/CRI, Birte Schrader ETH, PPM framework) - Implementation of adaptive multi-resolution particle methods (Ivo F. Sbalzarini ETH/CRI, Birte Schrader ETH) Besides delivering more accurate and individualized brain activity transmission matrices, the proposed project will also contribute to answering the following open questions: - What is the best hardware/software architecture to perform such simulations? - How can existing architectures be improved and further developed to meet the future needs in high-performance computing and networking? - What accurate direct transmission matrix is to be injected in real-time inverse problems for brain computing interfaces? - What will be the eventual customisation of these direct matrices according to different head/brain shapes? - How large are the inter-personal fluctuations in geometry, physical parameters, and transmission matrix? - What is the best way of validating models of human brain activity? - What are the most comfortable mental control commands? 4. Publication potential Due to its interdisciplinary nature (medicine, psychology, psychiatry, computational science, numerical mathematics, physics, engineering), we believe that the present project offers interesting publication opportunities. Given the broad interest in the topic, from hospitals over brain research to the telecommunication industry, we expect that more than one publication will appear in an interdisciplinary scientific journal of high profile and visibility. 6 to 8 other articles could be published in international journals of biophysics (material parameters and wave propagation dynamics in brain tissue, simulating and calibrating EEG and MEG measurements), computational science (adaptive multi-resolution particle method, parallel high-performance brain simulations), medicine (brain imaging, inter-person variations of brain parameters and geometries), psychiatry (brain imaging), and engineering (brain computer interface). Since the results and data produced are expected to find numerous applications beyond the scope of the present project, the publications bear the potential to be cited frequently. 5. Requested resources To complement the project team, we are looking for a full-time researcher on the postdoctoral level. The requested person will form the core of the proposed project, integrating and coordinating all the participating groups and researchers. As such, we are looking for a person that unites expertise in physics, applied mathematics, parallel computing, and biological/medical modeling. The proposed interdisciplinary post-doc will be able to discuss with specialists from all the different fields, gather geometric data of different head tropisms, formulate the physical models, and help implementing the simulation code using the PPM framework. He will be the principal person in modelling, performing the actual simulations, validating them, and discussing the results. 6. References [1] I. F. Sbalzarini, J. H. Walther, M. Bergdorf, S. E. Hieber, E. M. Kotsalis, and P. Koumoutsakos. PPM – a highly efficient parallel particle-mesh library for the simulation of continuum systems. J. Comput. Phys., 215(2):566–588, 2006. [2] M. Bergdorf, G.-H. Cottet, and P. Koumoutsakos. Multilevel adaptive particle methods for convection-diffusion equations. Multiscale Model. Simul., 4(1):328–357, 2005. [3] I. F. Sbalzarini, A. Mezzacasa, A. Helenius, and P. Koumoutsakos. Effects of organelle shape on fluorescence recovery after photobleaching. Biophys. J., 89(3):1482–1492, 2005. [4] N. Hansen and A. Ostermeier. Adapting arbitrary normal mutation distributions in evolution strategies: The covariance matrix adaptation. In Proceedings of the 1996 IEEE Conference on Evolutionary Computation (ICEC ’96), pages 312–317, 1996. [5] P. Koumoutsakos. Multiscale flow simulations using particles. Annu. Rev. Fluid Mech., 37:457–487, 2005. ____________________________________________________________________________________ Sponsored Link Rates near 39yr lows. $420,000 Loan for $1399/mo. Calcuate new payment. www.LowerMyBills.com/lre