Post on 25-Nov-2018
1
BrainScaleS FP7-ICT-2009 269921
Brain-inspired multiscale computation in neuromorphic hybrid systems
3rd Annual Meeting Institut de Neurosciences de la Timone
Marseille – France March 21 & 22, 2013
Abstracts
Report Version: 1.0 Classification: Consortium Type: Report Due date: Project month 18 Date issued: November 2012 Report Preparation: Contract Start Date: 1 January 2011 Duration: 4 years Project Coordinator: Karlheinz Meier (Universität Heidelberg, UHEI) Partners: Universitat Pompeu Fabra (UPF), Technische Universität
Dresden (TUD), Centre National de la Recherche Scientifique UNIC (Gif-sur-Yvette), INT and INS (Marseille), Technische Universität Graz (TUG), Forschungszentrum Jülich GmbH (JÜLICH), École Polytechnique Fédérale de Lausanne LCN (EPFL-LCN) and the Blue Brain Project (EPFL-BBP), Institut National de Recherche en Informatique et en Automatique (INRIA), Kungliga Tekniska Högskolan (KTH), Universität Zürich (UZH)
Additional partners since 1 Aug 2011: Koninklijke Nederlandse Akademie van Wetenschappen (KNAW), Universitetet For Miljo Og Biovitenskap (UMB), The Chancellor, Masters and Scholars of the University of Cambridge (CAM), Debreceni Egyetem (UD) and The University Of Manchester (UNIMAN)
2
DELIVERABLE SUMMARY SHEET
Project Number: 269921
Project Acronym: BrainScaleS
Deliverable N°:
Due date:
Delivery Date:
Short description:
Abstracts and supplementary material of the talks given at the 3d Annual Meeting of rhe BrainScaies project.
Partners owning: UHEI, CNRS-INT
Partners contributed: All
Made available to: Public
3
3rd BrainScaleS Plenary meeting - Agenda
Wednesday, 20 March 2013Wednesday, 20 March 2013Wednesday, 20 March 2013Wednesday, 20 March 201318:00-19:00 (60 min) Invited Lecture: Pr Ranulfo RomoInvited Lecture: Pr Ranulfo Romo
20:00-22:00(120 min)
Welcome dinnerat the Rowing Club Marseille34 Boulevard Charles Livon, 13007 Marseilletel: +33 (0)491 522715
Welcome dinnerat the Rowing Club Marseille34 Boulevard Charles Livon, 13007 Marseilletel: +33 (0)491 522715
Thursday, 21 March 2013Thursday, 21 March 2013Thursday, 21 March 2013Thursday, 21 March 2013Thursday, 21 March 201309:00 3rd BrainScaleS plenary meeting, day I3rd BrainScaleS plenary meeting, day I3rd BrainScaleS plenary meeting, day I3rd BrainScaleS plenary meeting, day I
09:00-09:15 (15 min) Welcome to MarseilleWelcome to Marseille Guillaume Masson (CNRS INT)
09:15-09:30 (15 min) Meeting introducationMeeting introducation Karlheinz Meier (UHEI)
09:30 WP 4: Development of Methods: Ideas, Tool & DevelopmentsWP 4: Development of Methods: Ideas, Tool & DevelopmentsWP 4: Development of Methods: Ideas, Tool & DevelopmentsWP 4: Development of Methods: Ideas, Tool & Developments09:30-09:40 (10 min) Introduction to the session: "Why methods and tools are the key to artificial brain-like systems"Introduction to the session: "Why methods and tools are the key to artificial brain-like systems" Laurent Perrinet (CNRS INT)
09:40-09:55 (15+5 min) Knowledge baseKnowledge base Andrew Davison (CNRS UNIC)
10:00-10:15 (15+5 min) Wave propagation in mice somatosensory Cortex: modelisation and parameters estimationWave propagation in mice somatosensory Cortex: modelisation and parameters estimation Nicolas Schmidt (Ceremade)
10:20-10:35 (15+5 min) Analysis of propagating waves from VSDI recordingsAnalysis of propagating waves from VSDI recordings Lyle Muller (CNRS UNIC)
10:40-11:10 (30 min) Coffee breakCoffee breakCoffee break
11:10-11:25 (15+5 min) Magnetrodes: seeking the magnetic field of neurons at the micron scaleMagnetrodes: seeking the magnetic field of neurons at the micron scale Myriam Pannetier-Lecoeur (CEA)
11:30-11:55 (25+5 min) How to do neuromorphic computing: from theory to experimentHow to do neuromorphic computing: from theory to experiment Mihai Petrovici (UHEI)
12:00-12:15 (15+5 min) MozaikMozaik Jan Antolik (CNRS UNIC)
12:20-12:25 (5 min) Group photoGroup photo
12:25-13:40 (75 min) LunchLunchLunch
13:40 Blue skyBlue skyBlue skyBlue sky13:40-13:55(15 min)
Blue sky: ideas, required input and necessary collaborations, next stepsBlue sky task leadersThe 'blue sky tasks' are:
• WP1 the Task 5: Blue Sky search for missing interaction terms between the different integration levels. Analogue signalling in axons. Validation of mean-field models (WP2 and WP4) and search for field effects. (leading partners: _CNRS-UNIC_ & _UPF_, contributing: KTH)
• WP4 the Task 7: Blue-sky methods: development of experimental and theoretical methods comprising high risk aspects, such as magnetic fields in neurons (_CNRS-UNIC_, CNRS-INCM)• WP7 the Task 6: Task 6: Solving variational problems and PDEs with large networks of spiking neurons (_INRIA_, UHEI, CNRS-UNIC, Jülich)
Blue sky: ideas, required input and necessary collaborations, next stepsBlue sky task leadersThe 'blue sky tasks' are:
• WP1 the Task 5: Blue Sky search for missing interaction terms between the different integration levels. Analogue signalling in axons. Validation of mean-field models (WP2 and WP4) and search for field effects. (leading partners: _CNRS-UNIC_ & _UPF_, contributing: KTH)
• WP4 the Task 7: Blue-sky methods: development of experimental and theoretical methods comprising high risk aspects, such as magnetic fields in neurons (_CNRS-UNIC_, CNRS-INCM)• WP7 the Task 6: Task 6: Solving variational problems and PDEs with large networks of spiking neurons (_INRIA_, UHEI, CNRS-UNIC, Jülich)
Karlheinz Meier (UHEI)
13:55 Demo 1 (WP5): Emerging representations & open-loop dynamics in large multi-scale models of sensory corticesDemo 1 (WP5): Emerging representations & open-loop dynamics in large multi-scale models of sensory corticesDemo 1 (WP5): Emerging representations & open-loop dynamics in large multi-scale models of sensory corticesDemo 1 (WP5): Emerging representations & open-loop dynamics in large multi-scale models of sensory cortices
13:55-14:05 (10 min) Demo 1 introducationDemo 1 introducation Anders Lansner (KTH)
14:05-14:20 (15+5 min) Decorrelating effects of inhibitory feedback in recurrent networksDecorrelating effects of inhibitory feedback in recurrent networks Markus Diesmann (Jülich)
14:25-14:40 (15+5 min) Compensation of hardware-specific distortions: models and methodsCompensation of hardware-specific distortions: models and methods Paul Müller (UHEI)
14:45-15:00 (15+5 min) Micro- and mesoscopic representation of apparent motion in S1Micro- and mesoscopic representation of apparent motion in S1 Dan Shulz (CNRS UNIC)
15:05-15:35 (30 min) Coffee breakCoffee breakCoffee break
15:35-15:50 (15+5 min) Demo 1 model : S1Demo 1 model : S1 Andrey Maximov (Jülich)
15:55-16:10 (15+5 min) A multi-scale approach to cortical representation of visual sceneryA multi-scale approach to cortical representation of visual scenery Björn Kampa (UZH)
16:15-16:30 (15+5 min) Demo1 model: Motion-based prediction in a network of spiking neuronsDemo1 model: Motion-based prediction in a network of spiking neurons Bernhard Kaplan (KTH)
16:35 Demo 3 (WP7): Beyond the realm of brain science: New directions in large-scale spike-based computingDemo 3 (WP7): Beyond the realm of brain science: New directions in large-scale spike-based computingDemo 3 (WP7): Beyond the realm of brain science: New directions in large-scale spike-based computingDemo 3 (WP7): Beyond the realm of brain science: New directions in large-scale spike-based computing
16:35-16:55 (20 min) Demo 3: IntroductionDemo 3: Introduction Wolfgang Maass (TUG)
16:55-17:10 (15+5 min) Neural noise and the stochastic properties of cortical neuronsNeural noise and the stochastic properties of cortical neurons Alain Destexhe (CNRS UNIC)
17:15-19:15(120 min)
Extended Steering Committee Meeting // Poster set up // Ethics Committee // Lab visitsLab visits:
• In vivo imaging (R-1)• Human Behavior (R+1)• In vitro electrophysiology (R+2) • Primate neurophysiology (R+4)
Extended Steering Committee Meeting // Poster set up // Ethics Committee // Lab visitsLab visits:
• In vivo imaging (R-1)• Human Behavior (R+1)• In vitro electrophysiology (R+2) • Primate neurophysiology (R+4)
19:30-22:00(150 min)
Poster-dinner (Location: INT ground floor)
The suggested poster theme is ... "demos":- for biologists to show experiments that can feed the demos- for modelers to show models that are supposed to be implemented in demos- for demo-implementers: closed loop simulation ideas / requirements / open issues- for software developers to show a complete simulation workflow including mapping and routing- for HW people: parameter space for hardware implementable models- for SpiNNaker: model implementations ideas
The poster session starts with a few sentences introduction (without slides) of each poster by the poster-presenter(s).
Registered posters:1. CNRS-UNIC: Michelle Rudolph and Lyle Muller "Aspects of randomness in biological neural graph structures"2. UMB: Gaute Einevoll: poster on the joint UMB-JULICH task on LFP modeling from spiking-network simulations3. Jülich:"Cortical multi-layered models for down-scaled implementation on neuromorphic hardware and full-scale implementation on supercomputers"4. UHEI Eric Müller and Paul Müller: A Closed-Loop Toy Experiment on Asynchronously Inter-Connected Compute Nodes5. UZH Poster by P. Molina-Luna, A.R. Woodruff, M.M. Roth, D.R. Muir, F. Helmchen and B.M. Kampa Sparse coding in neuronal subpopulations of mouse visual cortex during natural movie presentation6. UZH Poster by D.R. Muir, P. Molina-Luna, F. Helmchen and B.M. Kampa Specific connectivity and feature binding in mouse visual cortex7. UPF Etienne Hugues Sequential decision making8. UHEI Mihai el al. Neural sampling9. UHEI Mihai et al.Compensation of HW-specific distortions10. Jülich: de Haan MJ, Torre E, Zehl L, Denker M, Ito J, Brochier T, Grün S, Riehle A Massively parallel electrophysiological recordings from monkey primary visual and motor cortices (V1 and M1) during complex visually guided tracking tasks11. Jülich: Riehle A, Grün S, Brochier T Mapping the spatio-temporal structure of motor cortical LFPs related to reaching and grasping12. KTH: Phil Tully spiking implementation of BCPNN learning rule13. KTH: Pierre Berthet, Bernhard Kaplan demo2 (with parts and input from demo1)14. TUD Johannes Partzsch, Alex Rast, Bernhard Vogginger, Christian Mayr, Luis Plana, Stefan Schiefer, Mathias Ehrlich A prototype wafer-SpiNNaker communication demonstrator15. CAM: Autoassociative memories in neural networks with biological constraints16. UZH Simon Musall, Wolfger von der Behrens, Johannes Mayerhofer, Fritjof Helmchen und Florent HaissThe role of neural adaptation for tactile perception in primary somatosensory cortex17. UHEI Exploring the HICANN configuration space18. KTH Martin Rehn, David Silverstein and Anders Lanser: the KTH biophysical V1 model19. CNRS-INT: T Deneux, T Masquelier, G S. Masson, G Deco and I Vanzetta. The Spatiotemporal Structure of Ongoing and Evoked Activity investigated using Optical Imaging of Voltage Sensitive Dyes in Awake Monkey V4
Registered posters:1. CNRS-UNIC: Michelle Rudolph and Lyle Muller "Aspects of randomness in biological neural graph structures"2. UMB: Gaute Einevoll: poster on the joint UMB-JULICH task on LFP modeling from spiking-network simulations3. Jülich:"Cortical multi-layered models for down-scaled implementation on neuromorphic hardware and full-scale implementation on supercomputers"4. UHEI Eric Müller and Paul Müller: A Closed-Loop Toy Experiment on Asynchronously Inter-Connected Compute Nodes5. UZH Poster by P. Molina-Luna, A.R. Woodruff, M.M. Roth, D.R. Muir, F. Helmchen and B.M. Kampa Sparse coding in neuronal subpopulations of mouse visual cortex during natural movie presentation6. UZH Poster by D.R. Muir, P. Molina-Luna, F. Helmchen and B.M. Kampa Specific connectivity and feature binding in mouse visual cortex7. UPF Etienne Hugues Sequential decision making8. UHEI Mihai el al. Neural sampling9. UHEI Mihai et al.Compensation of HW-specific distortions10. Jülich: de Haan MJ, Torre E, Zehl L, Denker M, Ito J, Brochier T, Grün S, Riehle A Massively parallel electrophysiological recordings from monkey primary visual and motor cortices (V1 and M1) during complex visually guided tracking tasks11. Jülich: Riehle A, Grün S, Brochier T Mapping the spatio-temporal structure of motor cortical LFPs related to reaching and grasping12. KTH: Phil Tully spiking implementation of BCPNN learning rule13. KTH: Pierre Berthet, Bernhard Kaplan demo2 (with parts and input from demo1)14. TUD Johannes Partzsch, Alex Rast, Bernhard Vogginger, Christian Mayr, Luis Plana, Stefan Schiefer, Mathias Ehrlich A prototype wafer-SpiNNaker communication demonstrator15. CAM: Autoassociative memories in neural networks with biological constraints16. UZH Simon Musall, Wolfger von der Behrens, Johannes Mayerhofer, Fritjof Helmchen und Florent HaissThe role of neural adaptation for tactile perception in primary somatosensory cortex17. UHEI Exploring the HICANN configuration space18. KTH Martin Rehn, David Silverstein and Anders Lanser: the KTH biophysical V1 model19. CNRS-INT: T Deneux, T Masquelier, G S. Masson, G Deco and I Vanzetta. The Spatiotemporal Structure of Ongoing and Evoked Activity investigated using Optical Imaging of Voltage Sensitive Dyes in Awake Monkey V4
Friday, 22 March 2013Friday, 22 March 2013Friday, 22 March 2013Friday, 22 March 201308:45 BrainScaleS plenary meeting, day IIBrainScaleS plenary meeting, day IIBrainScaleS plenary meeting, day II
08:45-08:55 (10 min) Welcome to day II Guillaume Masson (CNRS INT)
08:55 3rd BrainScaleS plenary meeting, day II -- WP7 continued3rd BrainScaleS plenary meeting, day II -- WP7 continued3rd BrainScaleS plenary meeting, day II -- WP7 continued
08:55-09:10 (15+5 min) Variability vs. stability of neuronal responses measured with in vivo calcium imaging Fritjof Helmchen (UZH)
09:15-09:30 (15+5 min) Spike train statistics and Gibbs distribution Bruno Cessac (INRIA)
09:35-09:50 (15+5 min) Variability vs. stability of neuronal responses measured with in vivo calcium imaging Mate Lengyel (CAM)
09:55-10:05 (10 min) Theory of LIF neural sampling Ilja Bytschok (UHEI)
10:05-10:15 (10+5 min) Applications of LIF neural sampling Ilja Bytschok (UHEI)
10:20-10:50 (30 min) Coffee breakCoffee break
10:50-11:05 (15+5 min) Learning general probabilistic inference in networks of spiking neurons Dejan Pecevski (TUG)
11:10 Demo 2 (WP6): Multi-scale reconfiguration of cortical-like networks during active perceptionDemo 2 (WP6): Multi-scale reconfiguration of cortical-like networks during active perceptionDemo 2 (WP6): Multi-scale reconfiguration of cortical-like networks during active perception
11:10-11:20 (10 min) Demo 2: Introduction Guillaume Masson (CNRS INT)
11:20-11:35 (15+5 min) Feedforward and feedback processing for multiscale texture segregation Pieter Roelfsema (KNAW)
11:40-11:55 (15+5 min) How does anatomy shape dynamics in brain networks? Victor Jirsa (CNRS-ISM)
12:00-12:15 (15+5 min) Integrating multi-scale data for a network model of macaque visual cortex Sacha v. Albada (Jülich)
12:20-13:30 (70 min) LunchLunch
13:30-13:45 (15+5 min) Towards Closed-Loop Experiments on the Hybrid Multiscale Facility -- a preparatory study Eric Müller (UHEI)
13:50-14:05 (15+5 min) Demo2 model: Decision making in somatosensory system Etienne Hugues (UPF)
14:10-14:25 (15 min) Advisor commentsAdvisor comments Ulrich Rückert (Advisory board)
14:25-14:55 (30 min) Demos -- next stepsDemo WP leadersDemos -- next stepsDemo WP leaders
14:55-15:00 (5 min) Good byeGood bye Karlheinz Meier (UHEI)
15:00 End of the 3rd BrainScaleS plenary meetingEnd of the 3rd BrainScaleS plenary meetingEnd of the 3rd BrainScaleS plenary meeting
4
Special Lecture
Conversion of sensory signals into perceptual decisions Ranulfo Romo, IFC-‐UNAM, Mexico City, Mexico Most perceptual tasks require sequential steps to be carried out. This must be the case, for example, when subjects discriminate the difference in frequency between two mechanical vibrations applied sequentially to their fingertips. This perceptual task can be understood as a chain of neural operations: encoding the two consecutive stimulus frequencies, maintaining the first stimulus in working memory, comparing the second stimulus to the memory trace left by the first stimulus, and communicating the result of the comparison to the motor apparatus. Where and how in the brain are these cognitive operations executed? We addressed this problem by recording single neurons from several cortical areas while trained monkeys executed the vibrotactile discrimination task. We found that primary somatosensory cortex (S1) drives higher cortical areas where past and current sensory information are combined, such that a comparison of the two evolves into a decision. Consistent with this result, direct activation of the S1 can trigger quantifiable percepts in this task. These findings provide a fairly complete panorama of the neural dynamics that underlies the transformation of sensory information into an action and emphasize the importance of studying multiple cortical areas during the same behavioral task. Recommended article on which my talk will be based: Romo, R. & de Lafuente, V. Conversion of sensory signals into perceptual decisions. Progress in Neurobiology (2012)http://dx.org.org110.1016/j.pneurobio.2012.03.007
5
Session 1 – Tools, Ideas and Methods
BrainScaleS software toolchain: PyNN, Helmholtz and the Knowledge Base Andrew Davison and Domenico Guarino UNIC, CNRS Gif sur Yvette, France This talk will give an overview of the BrainScaleS software toolchain, and will then give updates on three of the components.
PyNN is a simulator-‐independent API for neuronal network modeling, and has been adopted as the standard model and simulation description format for BrainScaleS. We will present several new features that have been added to the development version, and that will be in the upcoming 0.8 release.
Helmholtz is a framework for building neuroscience databases, and was the basis for the VisionDB database in FACETS. We will present some important recent improvements, in particular an interface with the Elphy electrophysiology software and a web-‐services interface allowing programmatic database access from any programming language.
The BrainScaleS Knowledge Base attempts to systematically capture the knowledge that is shared between different labs and different workpackages. The most important pieces of knowledge will be encoded in machine-‐readable format and used for model building and model validation in the Demo workpackages. We will give a brief overview of how to use the Knowledge Base. Further readings & references PyNN; http://neuralensemble.org/PyNN/ Helmholtz : https://www.dbunic.cnrs-‐gif.fr/visiondb/ https://brainscales.kip.uni-‐heidelberg.de/jss/FileStore/dI_1873/BrainScaleS_D4-‐3.1.pdf Knowledge Base: https://www.dbunic.cnrs-‐gif.fr/knowledgebase/ https://brainscales.kip.uni-‐heidelberg.de/jss/FileStore/dI_1527/BrainScaleS_D5-‐1.2.pdf Waves Propagations in Mice Somatosensory Cortex: Models and Parameters Estimation. Nicolas Schmidt Ceremade, Université Paris-‐Dauphine Paris, France In this talk I will present a novel approach to model and extract the activity patterns in the mouse neocortex, observed with Voltage-‐sensitive dye optical imaging (VSDOI) [4]. The denoised VSDOI signal is modeled as the solution of a linear integro-‐differential equation. This model depends on a spatially localized source and a propagating/diffusive medium. This medium is defined though temporal linear filters that drive the geometry of the activity patterns. This model allows us to describe various phenomena observed in VSDOI imaging [3], such as time-‐frequency dissipation, propagation and diffusion. The parameters of the models provide meaningful information about the
6
observed patterns, such as the speed of propagation and the rate of diffusion. Recovering these medium parameters and the source from the VSDOI observations is a difficult inverse problem. A pre-‐processing step enables the detection of the sources spatial locations. We then estimate the filters parameterizing the medium though a variational optimization problem [1,2]. This is a joint work with Gabriel Peyré (CNRS and Univ. Paris-‐Dauphine), Yves Fregnac and Isabelle Ferezou (CNRS and UNIC). This work is supported by the ERC project SIGMA-‐Vision Further readings & references (http://www.ceremade.dauphine.fr/~peyre/sigma-‐vision/). [1] N. Schmidt, G. Peyré, Y. Fregnac, P. E. Roland, Separation of Traveling Waves in Cortical Networks
Using Optical Imaging, Proc. of ISBI'10, IEEE Press, pp. 868-‐871, 2010. [2] N. Schmidt, G. Peyré, Y. Fregnac, Dissipative Wave Model Fitting Using Localized Sources, Proc. Waves 2011, pp. 473-‐476, 2011. [3] Isabelle Ferezou, Sonia Bolea, Carl C.H. Petersen, Visualizing the Cortical Representation of
Whisker Touch: Voltage-‐Sensitive Dye Imaging in Freely Moving Mice, Neuron, Vol. 50(4), pp. 617-‐629, 2006
[4] A. Grinvald and R. Hildesheim. VSDI : A new era in functional imaging of cortical dynamics. Nature Reviews Neuroscience, 5(11), 2004.
Propagating waves during waking states: Discrimination and analysis Lyle Muller, UNIC, CNRS, Gif-‐sur-‐Yvette, France Propagating waves of activity are seen in many types of excitable media, and in recent years, were found in neural systems ranging from retina to neocortex. It remains unclear, however, whether waves appear during awake and conscious states. One possibility is that these waves are systematically missed in trial-‐averaged data, due to their variability. Here, using a phase-‐based analysis of single-‐trial voltage-‐sensitive dye imaging data, we show that spontaneous and stimulus-‐evoked propagating waves occur in visual cortex of the awake monkey. Furthermore, we observe correlated propagations across primary and secondary visual cortex, illustrating a strong spatiotemporal organization of these waves across cortical areas. These results suggest that propagating waves, systematically and reliably evoked by sensory stimulation, affect large-‐scale information processing by generating a consistent spatiotemporal frame for neuronal interactions. Further readings & references Grinvald et al. (1994): http://www.jneurosci.org/content/14/5/2545.short Xu et al. (2007): http://www.cell.com/neuron/abstract/S0896-‐6273(07)00446-‐1 Ray and Maunsell (2011): http://www.jneurosci.org/content/31/35/12674.long Reynaud et al. (2012): http://www.jneurosci.org/content/32/36/12558.short Muller and Destexhe (2012): http://www.sciencedirect.com/science/article/pii/S0928425712000393
7
MAGNETRODES: Seeking the magnetic field of neurons at the micron scale Myriam Pannetier-‐Lecoeur Service de physique de l’Etat Condensé, CEA Saclay France Spin electronics offers nowadays the possibility to create very sensitive, micrometer-‐scale magnetic field detectors. ‘Magnetrodes’ is an FP7-‐FET project, started in January 2013, aiming to exploit this technological advance to create novel tools for probing neuronal magnetic fields at the cellular level. The first goal of the project will be to develop the magnetic equivalent of an electrode, a ‘magnetrode’, sensitive enough to detect the very small magnetic fields induced by the ionic currents flowing within electrically active neurons, and small enough to probe a limited number of cells. We target also to adapt magnetrodes also for local nuclear magnetic resonance spectroscopy (MRS); thus, they could record both electromagnetic and chemical activity of neurons. In addition, means for local electric or magnetic stimulation could be integrated in to a magnetrode. We will test magnetrodes in vitro and in vivo at various spatial scales, from brain areas down to single neurons. In parallel, based on the measurements with magnetrodes, we will augment existing computational models and develop new ones to characterize the electromagnetic fields emitted by neurons and neuron assemblies. We will use these models to bridge from the activity of single neurons to macroscopic non-‐invasive measurements such as electroencephalography (EEG) and magnetoencephalography (MEG).
This project shall pave the way towards “magnetophysiology”, which enables investigating electric activity of neurons without disturbing the ionic flow and without physical contact to the cell. We will create new experimental and modeling tools for magnetic measurements and stimulation at neuron scale. The project consortium is composed of 5 teams from 4 EU countries.
How to do neuromorphic computing: from theory to experiment Mihai A. Petrovici, University of Heidelberg Heidelberg, Germany While saying that neuromorphic emulation is markedly different from software simulation might be stating the obvious, the past several years have brought a lot of understanding, both qualitative and quantitative, of the underlying reasons. With obvious advantages in speed and power consumption, the BrainScaleS neuromorphic hardware was designed as a universal emulator. However, neither the configurability nor the precision of such an analog device can match those of software simulators. In order to best exploit the advantages offered by neuromorphic circuits, the emerging breed of "neuromorphic modelers" needs to develop a novel mindset, both in terms of theory, as well as model design. Through intense collaborations within the consortia of both FACETS and BrainScaleS, various concepts have come forth, ranging from "software-‐simulation-‐related compensation of hardware-‐induced distortions" to "self-‐calibrating model implementations". We will review several of these techniques and thereby try to define a roadmap for future steps in the exploitation of neuromorphic hardware. Further readings & references http://www.kip.uni-‐heidelberg.de/Veroeffentlichungen/details.php?id=2766 http://www.kip.uni-‐heidelberg.de/Veroeffentlichungen/details.php?id=2185 http://www.kip.uni-‐heidelberg.de/Veroeffentlichungen/details.php?id=2018 https://brainscales.kip.uni-‐
8
heidelberg.de/internal/jss/AttendMeeting?m=displayPresentation&mI=46&mEID=860 https://brainscales.kip.uni-‐heidelberg.de/internal/jss/AttendMeeting?m=displayPresentation&mI=16&mEID=513 Mozaik: Integrated Workflow for advanced model specification and virtual experiment execution, analysis and visualization Jan Antolik UNIC, CNRS Gif-‐sur-‐Yvette, France This talk will introduce the mozaik framework. Mozaik is an integrated workflow framework, built on top of several tools used in FACETS and BrainScaleS (pyNN, NeuroToolS, Neo). It allows users to rapidly specify advanced heterogeneous models and subsequently test them in virtual experiments. To this end it allows user to easily specify the sensory and direct stimulation protocol and the recording configuration. It will automatically execute such experiments, collecting all the simulation results and link them to the appropriate meta-‐data from the experimental protocol. Furthermore it offers packages for analyzing and visualizing the recorded data, automatically using all the available meta-‐data, thus allowing for rapid specification of new analysis and visualization procedures, improving the productivity of the user. Currently mozaik has been mainly used in the domain of visual cortex modeling or for sensory-‐less models, but it should be applicable to other sensory modalities as well. Further readings & references https://github.com/antolikjan/mozaik.git
9
Session 2 – WP5 – DEMO 1
Decorrelating effects of inhibitory feedback in recurrent networks Markus Diesmann, Inst of Neuroscience and Medicine (INM-‐6), Computational and Systems Neuroscience & Institute for Advanced Simulation. Jülich Research Centre and JARA Jülich, Germany Depending on the researcher's favorite theory of cortical function, the correlation between the spike trains of neurons is seen as a limitation of the accuracy of neuronal processing, the carrier of the process, or just an epiphenomenon. Independent of the particular view held, it has been established that synaptic plasticity is governed by the relative timing of pre-‐ and postsynaptic activity, and hence by the correlation structure of the network.
In this talk I summarize our recent progress in understanding the correlation structure of neuronal activity resulting from basic assumptions about the architecture of the local cortical network and single neuron dynamics. We show that a negative feedback loop in the local network enables neurons in the asynchronous irregular state to be much less correlated than expected from the massive common input dictated by anatomy. Furthermore, the architecture prescribes a particular structure of the magnitude of correlations among the possible pairings of the neuron types [1] and the different shapes of the cross-‐correlation functions [2]. Further readings & references www.csn.fz-‐juelich.de www.nest-‐initiative.org [1] Tetzlaff T, Helias M, Einevoll GT, Diesmann M (2012) PLoS Comput Biol 8:e1002596 [2] Helias M, Tetzlaff T, Diesmann M (2013) New Journal of Physics 15(2):023002 Creating a toolset for neuromorphic modelers Paul Müller , Mihai A. Petrovici, Bernhard Vogginger & Oliver Breitwieser University of Heidelberg Heidelberg, Germany The core idea behind neuromorphic hardware is the emulation of neuronal dynamics, as opposed to their simulation on conventional computers. A physical implementation of neuronal circuits in silico offers significant advantages, especially with respect to emulation speed and power consumption. This approach is, however, not without its limitations, especially in terms of configurability, where conventional software excels by nature. In this respect, the BrainScaleS hardware[1] offers exceptional flexibility, allowing large freedom in network topology and parameter choices. With a software model of this device as a substrate, we have identified several important mechanisms which cause distortions in emulated network dynamics. In order to characterize the effects of these mechanisms, we have chosen three very different benchmark networks, in terms of both structure and dynamics. Our observations have allowed the design and implementation of various
10
compensation strategies. We hereby intend to offer a set of generic tools and concepts which should prove very useful in the interaction between modelers and neuromorphic hardware. Further readings & references [1] Realizing Biological Spiking Network Models in a Configurable Wafer-‐Scale Hardware System –
Johannes Schemmel, Johannes Fieres, Karlheinz Meier, Proceedings IJCNN2008, IEEE Press, 2008
[2] A Comprehensive Workflow for General-‐Purpose Neural Modeling with Highly Configurable Neuromorphic Hardware Systems – Daniel Brüderle, Mihai Petrovici, Bernhard Vogginger et al., Biological Cybernetics 104(4), 2011
[3] Bistable, Irregular Firing and Population Oscillations in a Modular Attractor Memory Network – Mikael Lundqvist, Albert Compte, Anders Lansner, PLoS Computational Biology 6(6), 2010
[4] Functional Consequences of Correlated Excitatory and Inhibitory Conductances in Cortical Networks, Jens Kremkow, Laurent Perrinet, Guillaume Masson, Ad Aertsen, Journal of Computational Neuroscience 28, 2010
[5] Self-‐Sustained Asynchronous Irregular States and Up/Down States in Thalamic, Cortical and Thalamocortical Networks of Nonlinear Integrate-‐and-‐Fire Neurons – Alain Destexhe, Journal of Computational Neuroscience 3, 2009
[6] Six networks on a universal neuromorphic computing substrate -‐ Thomas Pfeil, Andreas Grübl, Sebastian Jeltsch, Eric Müller, Paul Müller, Mihai A. Petrovici, Michael Schmuker, Daniel Brüderle, Johannes Schemmel, Karlheinz Meier, Front. Neurosci. 7(11), 2013
Micro-‐ and mesoscopic representation of apparent motion in S1 Shulz, D. E., Ego-‐Stengel, V., Vilarchao, E. & Férézou, I. UNIC, CNRS, Gif sur Yvette, France The tactile sensations mediated by the whisker-‐to-‐barrel cortex system allow rodents to efficiently detect and discriminate objects and surfaces. The temporal structure of whisker deflections and the temporal correlation between deflections occurring on several whiskers simultaneously vary for different tactile substrates. We hypothesize that tactile discrimination capabilities rely strongly on the ability of the system to encode different levels of inter-‐whisker correlations. To test this hypothesis, we generated complex spatio-‐temporal patterns of whisker deflections during electrophysiological single unit recordings in the barrel cortex, the ventro-‐posterior medial (VPM) nucleus of the thalamus and the trigeminal ganglion. A piezoelectric-‐based stimulator featuring 24 independent and fully adjustable whisker actuators was built for this purpose (Jacob et al., 2010). Using this stimulator in anesthetized rats, we have previously shown that cortical neurons exhibit direction selectivity to the apparent motion of a multivibrissal stimulus (i.e. an emerging property of the global stimulus), uncorrelated to the local direction of individual whiskers (Jacob et al. 2008). Since a certain level of multiwhisker integration has been reported in the VPM, the nucleus relaying tactile information to the barrel cortex, we showed that emergent properties of multiwhisker stimulations are already coded by VPM neurons although to a lesser degree than in cortex (Ego-‐Stengel et al., 2012).
We are presently exploring at a mesoscopic level the topographic representation of the selectivity to global motion in the barrel field of the primary somatosensory cortex using voltage sensitive dye imaging (VSD) in the anesthetized mouse. These data will be used to constrain a model of S1 that is being developed by Andrey Maximov and Sacha van Alba at Jülich as part of Demo 1 (WP5).
11
Supported by Facets-‐ITN (EV), ANR (IF) and HFSP (VES). Further readings & references http://www.unic.cnrs-‐gif.fr/site_media/pdf/Ego-‐Stengel2012.pdf http://www.unic.cnrs-‐gif.fr/site_media/pdf/voir_A695E2CFd01.pdf http://www.unic.cnrs-‐gif.fr/site_media/pdf/Jacob_Shulz08Neuron.pdf
Toward a large-‐scale spiking network model of the rodent barrel cortex Andrei Maksimov, Inst of Neuroscience and Medicine (INM-‐6), Computational and Systems Neuroscience & Institute for Advanced Simulation. Jülich Research Centre and JARA Jülich, Germany During the tactile exploration of the environment by the rodents, objects are contacted by whiskers generating complex spatiotemporal patterns of stimulations. This complex distributed information is analyzed by an array of corticocortical horizontal connections that provide, together with the multi-‐whisker thalamic input, a potential substrate for complex nonlinear temporal and spatial interactions. It was previously shown that barrel cortex neurons show selectivity to the global direction of an apparently moving tactile stimulus, suggesting that individual neurons combine and extract information from the entire whisker pad [1, 2]. To obtain a deeper understanding of the information processing in the rodent whisker system, we construct a realistic large-‐scale model of the barrel cortex. To this end, anatomical and electrophysiological data from a wide range of literature is integrated into a coherent whole. The model is implemented in NEST using the PyNEST interface [3] and consists of adaptive exponential integrate-‐and-‐fire neurons with conductance-‐based synapses and population-‐specific connection probabilities. First, a model of a single barrel column was developed, which includes key neuronal and network mechanisms, avoids parameter scaling, and yields layer-‐specific firing rates in realistic ranges. This will be further adjusted to reproduce realistic firing patterns and membrane potential dynamics during spontaneous and whisker-‐stimulated brain states. Features contributing to realistic dynamics include lognormally distributed synaptic strengths, voltage-‐dependent NMDA synapses, and the extension from point neurons to two or more compartments. Initial work on extending the model to multiple columns is presented. Further readings & references www.nest-‐initiative.org [1] Jacob V, Le Cam J, Ego-‐Stengel V, Shulz DE (2008) Emergent properties of tactile scenes selectively
activate barrel cortex neurons. Neuron 60: 1112-‐1125. [2] Estebanez L, El Boustani S, Destexhe A, Shulz DE (2012) Correlated input reveals coexisting coding
schemes in a sensory cortex. Nat Neurosci 15: 1691-‐1701. [3] Eppler JM, Helias M, Muller E, Diesmann M, Gewaltig M-‐O (2008) PyNEST: a convenient interface
to the NEST simulator. Front Neuroinformatics 2: 12.
12
A multi-‐scale approach to cortical representation of visual scenery Björn Kampa, Vision Lab Brain Research Institute, University of Zurich Zurich, Switzerland How is our visual environment represented and processed in the brain? In my lab, we seek answers to this fundamental question with a multi-‐scale approach combining two-‐photon imaging and electrophysiological recordings with computation model simulations. In this way, we can directly assess how neuronal response properties depend on the local network circuit. Connections between cortical neurons are not made randomly. Specific connections involving excitatory and inhibitory neurons have been measured both statistically and functionally in several areas of rodent neocortex. However, the precise composition of specific networks and the effect of specific connectivity on information processing in cortex remain in question, especially as a minority of synapses are likely to be made specifically. We found that specific excitatory connectivity can underlie amplification, decorrelation, competition and associative functions for cortex. Furthermore, our model simulations explain several observations of feature binding in visual cortex that we obtained using two-‐photon imaging of neuronal populations in mouse visual cortex. We also show that tuning for natural visual stimuli is independent of orientation preference, a likely consequence of specific connectivity. Our results suggest a population code, where the visual environment is dynamically represented in the activation of distinct functional sub-‐networks.
Motion-‐based prediction in a network of spiking neurons Bernhard Kaplan1 & Laurent Perrinet2 [1] Royal Institute of Technology, KTH, Sweden [2] Institut de Neurosciences de la Timone, CNRS/Aix-‐Marseille University The hypothesis of predictive coding, i.e. that the brain explicitly predicts future sensory input to establish a coherent representation of the world, is becoming generally accepted. However, it is not clear on which level neural networks implement such predictive coding and which function inhibitory neurons may have. Starting from an abstract framework which is based on the probabilistic representation of motion [1], we have developed a recurrent network model of conductance-‐based integrate-‐and-‐fire neurons inspired by the architecture of retinotopic cortical areas which assumes that the basis for predictive coding is implemented through network connectivity. We show that the applied network connectivity, which is based on the tuning properties of source and target cells, leads to motion-‐based prediction in a moving dot tracking experiment. In contrast to our proposed connection pattern, networks with isotropic (non-‐selective) or random connectivity fail to predict the trajectory when the moving dot disappears. The model is implemented in PyNN and is one suitable candidate to be run on the BrainScaleS HMF and could serve as input for a oculor-‐motor reward learning model as part of Demo 2. Further readings & references [1] Laurent U. Perrinet and Guillaume S. Masson Motion-‐Based Prediction Is Sufficient to Solve the
Aperture Problem. Neural Computation 24, 2726–2750 (2012)
13
Session 3 – WP7 – DEMO 3 Introduction Wolfgang Maass Institute for Theoretical Computer Science Graz University of Technology Graz, Austria I will sketch the goals of Demo 3, and our new perspectives for the coming years. In addition I will highlight the common theme of the presentations for Demo 3: The role of noise in spike-‐based computations. Finally I will introduce a new class of computational problems (constraint satisfaction problems) that is -‐-‐according to theory-‐-‐ within reach of spike-‐based hardware.
Neuronal noise and the stochastic properties of cortical neurons Alain Destexhe, UNIC, CNRS, Gif sur Yvette, France Neurons are subject to different noise sources, the largest being the irregular and seemingly stochastic activity of the network which causes intense and very noisy synaptic bombardment in single neurons ("synaptic noise"). At the single cell level, this bombardment is responsible for setting neurons into a stochastic mode of firing. In such a regime, computing spike probabilities is the right measure to monitor neural responses, as routinely done in vivo trough the use of post-‐stimulus time histograms (PSTH).
Models and experiments have shown that synaptic noise is responsible for several interesting properties, such as drastically changing the transfer function of neurons, which takes the form of "gain modulation". Interestingly, synaptic noise can enhance the responsiveness to small inputs, by mechanisms analogous to stochastic resonance. When simulated in dendritic trees, these properties may also change the properties of dendritic integration, and temporal resolution of the neuron. All these properties require that the "noise" is conductance-‐based, and is the result of a balance between excitatory and inhibitory inputs. At the network level, models can generate states of activity similar to in vivo recordings, but again they must be conductance-‐based and include excitatory and inhibitory neurons forming balanced states. We show that network states consistent with conductance measurements are possible and can be identified using mean-‐field formalisms.
Finally, we show that stochastic states of activity may show properties of enhanced information transfer in networks. Collectively, these results show that stochastic states can confer interesting computational properties to neurons. Further readings & references http://cns.iaf.cnrs-‐gif.fr/abstracts/Sci2006.html http://cns.iaf.cnrs-‐gif.fr/abstracts/NeuroComp2007.html http://cns.iaf.cnrs-‐gif.fr/abstracts/TCX2008.html http://cns.iaf.cnrs-‐gif.fr/abstracts/Master2008.html http://cns.iaf.cnrs-‐gif.fr/abstracts/CurrOpinion2011.html
14
Spike train statistics and Gibbs distributions Bruno Cessac, INRIA NeuroMathComp Sophia-‐Antipolis, France With the advent of new Multi-‐Electrod Arrays (MEA) techniques, the simultaneously recording of the activity of groups of neurons (up to several hundreds) over a dense configuration, supplies today a critical database to understand how information is encoded in the brain. However, beyond the acquisition of such (massive) data there is a need to develop suitable statistical models for data analysis. In this talk we shall argue that Gibbs distributions, considered in more general setting than the initial concept coming from statistical physics and thermodynamics (including non stationarity), are canonical models for spike train statistics analysis.
This statement is based on three examples briefly discussed in the talk: (1) Maximum entropy models; (2) Linear-‐ Non-‐linear and Generalized Linear Model (3) Exact results on spike statistics in conductance based Integrate and Fire models with chemical and electric synapses. The interest of Gibbs distributions is not only to provide a general setting to properly consider spatio temporal correlations in spike trains, it also offers an efficient tool to generate artificial spike trains with prescribed spatio-‐temporal correlations structure, that mimics real spike trains. Further readings & References http://lanl.arxiv.org/abs/1302.5007 http://lanl.arxiv.org/abs/1212.3577 ftp://ftp-‐sop.inria.fr/neuromathcomp/team/bruno.cessac/Papers/author.pdf http://www.sciencedirect.com/science/article/pii/S0928425711000441 http://www.mathematical-‐neuroscience.com/content/1/1/8 Variability vs. stability of neuronal responses measured with in vivo calcium imaging Fritjof Helmchen, Brain Research Institute, UZH, Zurich, Switzerland Optical measurements of neuronal network dynamics in the brain of living animals have become possible by combining novel techniques for labelling neuronal populations with fluorescent calcium indicators and two-‐photon laser-‐scanning microscopy. I will present recent progress in the application of genetically encoded calcium indicators (GECIs) to functionally probe neuronal populations in mouse neocortex, especially during behavior. We employ adeno-‐associated viral (AAV) constructs to express GECIs in mouse neocortex to follow neuronal activity in the exact same neurons over weeks and months. In particular, we are measuring neocortical network dynamics in awake, behaving mice adapted to tolerate head-‐fixation. Imaging results from the primary somatosensory cortex (S1) demonstrate broad response distributions with a tail of salient neurons
15
showing especially large responses. Analysis of whisking-‐related neuronal activity exhibited differential response types, which remained stable over weeks. Using post hoc immunostaining we are analyzing whether different response patterns relate to neuronal subtype, e.g. GABAergic neurons. Furthermore we collected S1 population activity data during a reward-‐based Go-‐Nogo texture discrimination task, for which we can analyze trial-‐to-‐trial variability of behavior-‐related activity in trained expert mice. Theory for LIF sampling Ilja Bytschok University of Heidelberg Heidelberg, Germany A characteristic feature of information processing in the brain is its robustness and efficiency in coping with noisy, multisensory data. Experimental evidence suggests that cortical reasoning and decision-‐making might incorporate stochastic inference algorithms. Applications for probabilistic computing are manifold, but they are usually resource-‐intensive, especially when the embedded algorithms run on intrinsically sequential conventional computing architectures. Büsing et al. [1] have shown how a network of stochastically spiking neurons can implement Gibbs sampling from Boltzmann distributions, thereby allowing inference to be performed by an inherently parallel substrate. We have transferred this framework to networks of deterministic LIF units, with stochasticity provided by diffuse background stimulation. Under such stimulation, the LIF neurons enter the high-‐conductance state, whose underlying dynamics can be interpreted as an Ornstein-‐Uhlenbeck process. We developed a method to compute the firing activity of a single LIF neuron by solving the First-‐passage-‐time problem [2,3] for finite synaptic time constants. Having understood the behavior of a single computing LIF unit, the network dynamics from the theoretical model in [1] can be transferred to the LIF domain by compensating for the rectangular PSP shapes in the theoretical model. The presented approach provides biological plausibility and allows straightforward implementation on the BrainScaleS hardware. Further readings & References [1] Buesing, Bill, Nessler, Maass. Neural Dynamics as Sampling: A Model for Stochastic Computation
in Recurrent Networks of Spiking Neurons. PLoS Computational Biology, 2011 [2] Thomas. Some Mean First-‐Passage Time Approximations for the Ornstein-‐Uhlenbeck Process.
Journal of Applied Probability, 1975 [3] Finch. Mathematical Constants. Cambridge University Press, 2004 Applications for LIF sampling Ilja Bytschok University of Heidelberg Heidelberg, Germany To explore the full computational potential of the Neural sampling paradigm proposed in [1], the LIF sampling network architectures require an inherently parallel substrate, such as the BrainScaleS neuromorphic hardware device [2]. We evaluated the sampling quality for extensive parameter
16
ranges and developed a software framework to account for potential hardware parameter mismatches when generating sampling networks on the hardware. These preparatory studies are crucial for the implementation of a broad range of probabilistic models based on the LIF sampling functionality. Such potential models are stochastic Winner-‐Take-‐All (WTA) modules, Bayesian networks [3] and Deep Boltzmann Machines [4], incorporating thousands of stochastic LIF units. But also entirely beyond the realm of brain science, we uncovered a link between LIF sampling networks and Ising models of Two-‐state systems. More precisely, LIF networks can sample from Boltzmann distributions which represent the probabilities of occuring configurations of spin systems in magnetic fields. This analogy allows us to reproduce well-‐known magnetic phenomena in our sampling system, such as hysteresis, phase transitions and Weiss domains. Such large stochastic systems require poisson background noise sources to operate in a stochastic regime. The hardware device itself cannot provide enough noise sources for each sampling neuron. Since shared noise sources would distort sampling dynamics through uncompensated correlations, we develop so-‐called Sea of noise networks to generate uncorrelated background input [5] and ensure high sampling quality. Further readings & References [1] Buesing, Bill, Nessler, Maass. Neural Dynamics as Sampling: A Model for Stochastic Computation
in Recurrent Networks of Spiking Neurons. PLoS Computational Biology, 2011 [2] Realizing Biological Spiking Network Models in a Configurable Wafer-‐Scale Hardware System -‐
Johannes Schemmel, Johannes Fieres, Karlheinz Meier, Proceedings IJCNN2008, IEEE Press, 2008
[3] Probabilistic inference in general graphical models through sampling in stochastic networks of spiking neurons -‐ Pecevski, D., Buesing, L. and Maass, W. PLoS Computational Biology (2011) 7(12)
[4] Deep Boltzmann Machines -‐ R. Salakhutdinov, G. Hinton, Artificial Intelligence and Statistics 200(2012)
[5] Decorrelation of neural-‐network activity by inhibitory feedback -‐ T. Tetzlaff, M. Helias, G.T. Einevoll, M. Diesmann, PLoS Comp Biol 8(8) (2012)
Learning probabilistic inference in general graphical models with networks of spiking neurons Dejan Pecevski Institute for Theoretical Computer Science Graz University of Technology Graz, Austria Many behavioral data as well as recent data in neuroscience suggest that the brain stores knowledge in form of probability distributions, which are used to make inferences about the world based on observed facts. But it is still largely unknown how plasticity processes on a synaptic and neuronal level in the brain enable learning of knowledge in from of probability distributions. Recent theoretical results [Buesing et al., 2011, Pecevski et al., 2011] showed a novel way how a network of stochastic spiking neurons can "embody" a probability distribution and perform probabilistic inference in it via Markov chain Monte Carlo sampling. We show in this work that a particular form of synaptic plasticity together with plasticity of the intrinsic neuron excitabilities derived from basic theoretical principles, enable learning of a probability distribution in such networks of spiking neurons from presented data samples. The approach can be applied for learning any probability
17
distribution represented by any graphical model structure. By having a connectivity structure that reflects the independencies in the graphical model the neural networks exploit this independencies to reduce the complexity of learning. We demonstrate the viability of the approach in two computer simulation examples, where we train neural networks to learn probabilistic models for two perceptual phenomena: perceptual explaining away and localization bias in multisensory integration. Further readings & References Pecevski D, Buesing L, Maass W (2011) Probabilistic Inference in General Graphical Models through
Sampling in Stochastic Networks of Spiking Neurons. PLoS Comput Biol 7(12): e1002294. Buesing L, Bill J, Nessler B, Maass W (2011) Neural Dynamics as Sampling: A Model for Stochastic
Computation in Recurrent Networks of Spiking Neurons. PLoS Comput Biol 7(11): e1002211.
18
Session 4 – WP6 – DEMO 2 Feedforward and feedback processing for multiscale texture segregation Pieter Roelfsema Department of Vision and Cognition, Netherlands Institute for Neuroscience, Royal Netherlands Academy of Arts and Sciences, Amsterdam, The Netherlands. Our visual system segments images into objects and background. Figure-‐ground segregation relies on the detection of feature discontinuities that signal boundaries between the figures and the background and on a complementary region-‐filling process that groups together image regions with similar features. The neuronal mechanisms for these processes are not well understood and it is unknown how they depend on visual attention. We measured neuronal activity in V1 and V4 in a task where monkeys either made an eye movement to texture-‐defined figures or ignored them. We found that boundary detection is an early process that depends little on attention, whereas region filling occurs later and is facilitated by visual attention, which acts in an object-‐based manner. Our findings are explained by a model with local, bottom-‐up computations for boundary detection and feedback processing for region filling (Poort et al., Neuron, 2012).
In addition, we investigated low frequency (alpha) and high-‐frequency (gamma) oscillations and found that they characterize different directions of information flow in monkey visual cortex. Alpha oscillations index feedback effects and gamma oscillations signal feedforward processing (van Kerkoerle et al., under revision). Thus, our results also provide new insights into the relation between brain rhythms and cognition. Further readings & References Wanig A, Stanisor L & Roelfsema P (2011) Automatic spread of attentional response modulation
along Gestalt criteria in primary visual cortex. Nature Neuroscience 14, 1243-‐1245 Poort J, Raudies F, Waning A, Lamme VAF, Neumann H & Roelfsema PR (2012) The role of attention
in figure-‐ground segregation in areas V1 and V4 of the visual cortex. Neuron, 75, 146-‐153 Roelfsema PR, Lamme VA, Spekreijse H & Bosch H (2002) Figure-‐ground segregation in a recurrent
network artchitecture. J Cogn Neurosci 14, 525-‐537 Roelfsema PR (2006) Cortical algorithms for perceptual grouping. Annual Reviews in Neuroscience,
29, 203-‐227 Integrating multi-‐scale data for a network model of macaque visual cortex Maximilian Schmidt Inst of Neuroscience and Medicine (INM-‐6), Computational and Systems Neuroscience & Institute for Advanced Simulation. Jülich Research Centre and JARA Jülich, Germany Models of cortical dynamics usually either cover small cortical circuits in detail, or represent large patches in a highly simplified manner, for instance using a few differential equations for each area. This is due partly to limited computational resources, and partly to sparsity of large-‐scale structural connectivity data. Making use of advances in NEST, supercomputing resources, and the increased availability of large-‐scale connectivity data, we construct a spiking network model of the 32 areas of the macaque cortex associated with visual processing [1]. The individual areas are based on our recent layered cortical microcircuit model [2]. The extension to multiple areas enhances the self
19
consistency of the model. A detailed connectivity map is derived using realistic area-‐ and layer specific neuron densities, in-‐degrees, and laminar thicknesses based on the experimental literature. Both local and inter-‐area connectivity are layer-‐specific. The CoCoMac database [3] provides the basis for the inter-‐area connectivity, and is supplemented with quantitative information from further tracing studies [4,5]. We complete these data sets by deriving novel empirical connectivity rules, for instance exploiting the dependence of connection strengths on inter-‐area distances, and of laminar connection patterns on architectural type differences [6]. The model was implemented in NEST with the data integration performed using Python, and simulation results are tracked using Sumatra [7]. Preliminary results show area-‐ and population-‐specific firing rates and degrees of irregularity. In a next step, resting-‐state networks will be investigated in collaboration with UPF, which will be facilitated by the development of a mean-‐field approximation to the model. Further readings & References www.nest-‐initiative.org [1] Schmidt M, van Albada S, Bakker R, Diesmann M. Toward a spiking multi-‐area network model of
macaque visual cortex. NWG 2013. [2] Potjans T, Diesmann M (2012) The cell-‐type specific cortical microcircuit: relating structure and
activity in a full-‐scale spiking network model. Cereb Cortex doi:10.1093/cercor/bhs358 [3] Stephan KE, Kamper L, Bozkurt A, Burns GAPC, Young MP, Kötter R (2001) Advanced database
methodology for the collation of connectivity data on the macaque brain (CoCoMac). Phil Trans R Soc Lond B, 356: 1159-‐1186.
[4] Markov NT, Misery P, Falchier A, Lamy C, Vezoli J et al. (2011) Weight consistency specifies regularities of macaque cortical networks. Cerebral Cortex 21: 1254-‐1272.
[5] Markov NT, Ercsey-‐Ravasz MM, Ribeiro Gomes AR, Lamy C, Magrou L et al. (2012) A weighted and directed interareal connectivity matrix for macaque cerebral cortex. Cereb Cortex doi:10.1093/cercor/bhs270
[6] Hilgetag CC, Grant S (2012) Cytoarchitectural differences are a key determinant of laminar projection origins in the visual cortex. NeuroImage 51: 1006.
[7] Davison AP (2012) Automated capture of experiment context for easier reproducibility in computational research. Computing in Science and Engineering 14: 48-‐56.
How does anatomy shape dynamics in brain networks? Viktor Jirsa Institut de Neurosciences des Systèmes, INSERM Marseille, France We use mathematical modeling and simulations to explore the dynamics that emerge in large scale cortical networks, with a particular focus on the topological properties of the structural connectivity and its relationship to functional connectivity. In particular we exploit realistic anatomical connectivity matrices (from diffusion spectrum imaging) and investigate their capacity to generate various types of resting state activity. The emergent patterns of activity for realistic connectivity configurations together with approximations are formulated in terms of neural mass or field models. We find that homogenous connectivity matrices, of the sort of assumed in certain neural field models give rise to damped spatially periodic modes, while more localized modes reflect heterogeneous coupling topologies. When simulating resting state fluctuations under realistic
20
connectivity, we find no evidence for a spectrum of spatially periodic patterns, even when grouping together cortical nodes into communities, using graph theory. We conclude that neural field models with translationaly invariant connectivity may be best applied at the mesoscopic scale and that more general models of cortical networks that embed local neural fields, may provide appropriate models of macroscopic cortical dynamics over the whole brain. Further readings & References http://thevirtualbrain.org/team/index.html http://ins.medecine.univmed.fr/fr/research-‐teams/theoretical-‐neurosciences-‐group/ Pinotsis DA, Hansen E, Friston KJ & Jirsa VK (2013) Anatomical connectivity and the resting state activity of large cortical networks. NeuroImage, 65: 127-‐138 Towards Closed-‐Loop Experiments on the Hybrid Multiscale Facility -‐-‐ a preparatory study Eric Müller, UHEI, Universität Heidelberg Heidelberg, Germany Closed-‐loop operation of neuromorphic hardware is a demanding task that poses hard constraints on all computational and communication components. In particular, to keep software in sync with hardware operating in continuous time it is essential to minimize latencies induced by communication and conventional computation. To test the performance of our conventional hardware setup and to quantify limitations of software involved in closed-‐loop experiments, we implemented a proof of concept experiment that already includes many characteristics of a full closed-‐loop experiment.
Two compute nodes of the Hybrid Multiscale Facility (HMF) cluster simultaneously run different parts of this experiment interacting asynchronously via spikes in real-‐time. One part simulates the "physical environment" while the other part simulates a spiking neural network and therefore acts as a surrogate for the HMF neuromorphic system. The interaction between physical environment and neural network is evaluated in terms of computational runtime and communication latency.
In addition, we present latency measurements of inter-‐node communication within the HMF conventional cluster, a method to synchronize cluster nodes down to 500ns and key figures for closed-‐loop modeling on the HMF. Demo 2 model: Decision making in somatosensory system Etienne Hugues, UPF, Barcelona, Spain Perceptual discrimination may be interpreted as a decision between alternatives based on available sensory evidence. In many experiments, the different alternatives are encoded by quite distinct neuronal groups. In this case, proposed neural models consider that the decision results from the
21
competition between decision-‐specific neuronal groups, each of these integrating distinct sensory evidence [1]. Alternatively, as in the vibrotactile discrimination task of Romo and collaborators [2], evidence may be presented in a sequential manner, and the different stimuli may be encoded by the same neuronal group. Beyond the fact that a memory of the previously presented stimulus should be kept, how the nervous system is able to accomplish the correct discrimination is poorly understood. In the vibrotactile discrimination task, the partial differential (PD) neurons in monkey area VPC, encoding both sequentially presented vibrotactile stimuli (with frequencies f1 and f2) by keeping the memory of the first one during a delay period, have been reproduced in a spiking neuron network mode,l where short-‐term facilitating synapses support the memory [3]. We want now to explore how these PD neurons may be used to discriminate between both stimuli configurations: f1 > f2 or f1 < f2. Based on the experimental evidence, we model a heterogeneous PD neuronal population, encoding both frequencies in multiple ways. Downstream to the first network, we add a competition-‐based decision making spiking neuron network [1]. To make the best possible decisions, the strengths of the synapses projecting from the PD neurons to the decision neurons must be learned. Based on reinforcement learning theory, we use a learning rule which maximizes reward, and depends on a reward prediction error which is evaluated using the reward history. Learning occurs after the second stimulus presentation. This rule can be instantiated using a reward based spike-‐timing-‐dependent plasticity [4]. We find that the task can be efficiently learned for any number of PD neuron types, even when their stimulus encoding function is nonlinear and noisy. With learning, the present biophysical two-‐networks model solves the sequential discrimination task in a closed loop manner. I will report here on the progress towards the full implementation of the model. Further readings & References 1. Wang X-‐J (2002) Probabilistic decision making by slow reverberation in cortical circuits, Neuron 36:955-‐ 968. http://www.cell.com/neuron/abstract/S0896-‐6273(02)01092-‐9 2. Hernández A, Nácher V, Luna R, Zainos A, Lemus L, Alvarez M, Vázquez Y, Camarillo L, Romo R (2010) Decoding a perceptual decision process across cortex, Neuron 66:300-‐314. http://www.cell.com/neuron/abstract/S0896-‐6273(10)00234-‐5 3. Deco G, Rolls ET, Romo R (2010) Synaptic dynamics and decision making. Proc. Natl Acad. Sci. USA 107:7545-‐7549. http://www.pnas.org/content/107/16/7545.abstract?sid=eeac4a62-‐da7f-‐4669-‐b27e-‐6d23c7a971b6 4. Frémeaux N, Sprekeler H, Gerstner W (2010) Functional requirements for reward-‐modulated spike-‐timing dependent plasticity. J. Neurosci 30(40):13326-‐13337. http://www.jneurosci.org/content/30/40/13326.abstract?sid=625ba64d-‐aaea-‐4590-‐bcf9-‐3632ce27a42d