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