Neural Simulation Language NSL

1. Neural Simulation LanguageNSL

2. Overview • Introduction • NSLM • Example (Max Selector) • NSLS • Downloading and installing NSL

3. Introduction • NSL is a platform for • Building neural architectures (modeling) • NSLM • NSLJ & NSLC • Executing them (simulation). • NSLS • NSL provides tools for modeling complex neural systems - especially (but not only) when the neurons are modeled as leaky integrator neurons.

4. Methodology • The general methodology for making a complex neural model of brain function is to combine different modules corresponding to different brain regions. • To model a particular brain region, we divide it anatomically or physiologically into different neural arrays. • Each brain region is then modeled as a set of neuron arrays, where each neuron is described for example by the leaky integrator, a single-compartment model of membrane potential and firing rate.

5. Levels of abstraction • A complete model in NSL requires the following components: • a set of modules defining the entire model • neurons comprised in each neural module • neural interconnections • neural dynamics • numerical methods to solve the differential equations.

6. Example of modules

7. Leaky integrator

8. Simulation 1 2 6 7 3 4 5

9. Simulation Methods • initSys • initModule • makeConn • … (simulation steps) • endModule • endSys

10. Simulation Methods • initTrainEpoch • initTrain • simTrain • endTrain • endTrainEpoch • initRunEpoch • initRun • simRun • endRun • endRunEpoch train doTrainEpochTimes trainAndRunAll run doRunEpochTimes

11. Model Structures • Model • Highest level • Modules • NeuralNetworks • InModules • OutModules • Graphic Interfaces • MotorModules • Robotics • NslClass • Libraries • New Canvases • New NSLS Commands

12. Primitive types int float double boolean char NslData types (0, 1, 2, 3, 4) NslInt NslFloat NslDouble NslBoolean NslString (0) Could be public, private or protected. NslPort types (0, 1, 2, 3, 4) NslDinInt NslDinFloat NslDinDouble NslDinBoolean NslDinString (0) NslDoutInt NslDoutFloat NslDoutDouble NslDoutBoolean NslDoutString (0) Ports must be public NSLM Types

13. Max Selector Model • The details of this model can be found in section 4.4 of TMB2. The model uses competition mechanisms to obtain, in many cases, a single winner in the network where the input signal with the greatest strength is propagated along to the output of the network.

14. Max Selector Model (2)

15. Max Selector Model (3) MaxSelectorModel MaxSelector Output Stimulus ULayer VLayer

16. MaxSelectorModel • nslImport nslAllImports; • nslImport MaxSelectorStimulus; • nslImport MaxSelector; • nslImport MaxSelectorOutput; • nslModel MaxSelectorModel() { • nslConst int size = 10; • private MaxSelectorStimulus stimulus(size); • private MaxSelector maxselector(size); • private MaxSelectorOutput output(size); • public void initSys() { • system.setRunEndTime(10.0); • system.nslSetRunDelta(0.1); • } • public void makeConn() { • nslConnect(stimulus.s_out, maxselector.in); • nslConnect(stimulus.s_out, output.s_out); • nslConnect(maxselector.out, output.uf); • } • }

17. MaxSelectorStimulus • nslImport nslAllImports; • nslModule MaxSelectorStimulus(int size) { • public NslDoutDouble1 s_out(size); • public void initRun() { • s_out=0; • s_out[1]=0.5; • s_out[3]=1.0; • } • }

18. MaxSelectorOutput • nslImport nslAllImports; • nslOutModule MaxSelectorOutput(int size) { • public NslDinDouble1 s_out(size); • public NslDinDouble1 uf(size); • private NslDouble1 up(size); • private boolean worked= false; • public void initModule() { • up.nslSetAccess('W'); • nslAddAreaCanvas(s_out,0,1); • nslAddTemporalCanvas(up,-2.5,2.5); • nslAddAreaCanvas(uf,0,1); • } • public void simRun() { • worked=nslSetValue(up,"maxSelectorModel.maxselector.u1.up"); • } • }

19. MaxSelector • nslImport nslAllImports; • nslImport Ulayer; • nslImport Vlayer; • nslModule MaxSelector(int size) { • public NslDinDouble1 in(size); • public NslDoutDouble1 out(size); • private Ulayer u1(size); • private Vlayer v1(); • public void makeConn() { • nslRelabel(this.in, u1.s_in); • nslConnect(u1.uf, v1.u_in); • nslConnect(v1.vf, u1.v_in); • nslRelabel(u1.uf, this.out); • } • }

20. ULayer • nslImport nslAllImports; • nslModule Ulayer(int size) { • //inports • public NslDinDouble1 s_in(); • public NslDinDouble0 v_in(); • //outports • public NslDoutDouble1 uf(size); • //variables • private NslDouble1 up(size); • private NslDouble0 w1(); • private NslDouble0 w2(); • private NslDouble0 h1(); • private NslDouble0 k(); • private double tau; • …

21. Ulayer(2) • public void initRun(){ • uf = 0; • up = 0; • tau = 1.0; • w1= 1.0; • w2= 1.0; • h1= 0.1; • k= 0.1; • } • public void simRun(){ • //compute : up=up+((timestep/tu)*du/dt) • up = nslDiff(up, tau, -up + w1*uf-w2*v_in - h1 + s_in); • uf = nslStep(up,k.get(),0,1.0); • } • }

22. VLayer • nslImport nslAllImports; • nslModule Vlayer() { • // ports • public NslDinDouble1 u_in(); • // output port • public NslDoutDouble0 vf(); • // variables • private NslDouble0 vp(); // neuron potential • private NslDouble0 h2(); • private double tau; // time constant • …

23. Vlayer (2) • public void initRun() { • vf=0; • vp=0; • tau=1.0; • h2 = 0.5; • } • public void simRun() { • // vp=vp+((timestep/tv)*dv/dt) • vp = nslDiff(vp, tau, -vp + nslSum(u_in) -h2); • vf = nslRamp(vp); • } • }

24. Compilation • One model/module per file • The file extension must be .mod • We recommend to clean the model directory before compiling with the nslclean command • To compile the model you just have to execute the following command: nslc modelName • Where modelName is the Name of the file that contains the model structure. • For this example we should write: nslc MaxSelectorModel • Note that we didn’t write the file extension at the end of the name. Class File Mod File Nlx File Java File

25. Execution • To simulate your model you have to use the nsl command. • For this example you should write: nsl MaxSelectorModel • Two running modes • Text (-nodisplay) • Graphic interface (default) • To redirect the standard output (-stdout console) • To redirect the standard error (-stderr console)

26. Interface

27. Interface (2)

28. NSLS • To avoid re-compiling every time you modify your model parameters we provide the NSL script language known as NSLS which also provides a dynamic user control environment. • NSLS provides the following functionality: • NSL model parameter assignment • NSL input specification • NSL simulation control • NSL file control • NSL graphics control • NSLS is an extension of the well know TCL scripting language, thus providing NSL and TCL functionality.

29. NSLS (2) • NSL command syntax: nsl subcommand [options] • Important NSL commands: • nsl source fileName • (i.e. nsl source hopfield.nsls) • nsl set variable value • (i.e. nsl set maxSelectorModel.stimulus.s_out {1 0 0.5}) • nsl get variable • (i.e. nsl get maxSelectorModel.stimulus.s_out) • nsl run • nsl train • … • nsl exit

30. NSLS example • # • # Hopfield Network • # • set A {} • set B {} • set C {} • set D {} • proc memorize { x } { • puts "Memorizing $x" • nsl set hopfield.inModule.input $x • nsl train • } • proc test { x d } { • nsl set hopfield.dis $d • nsl set hopfield.inModule.input $x • nsl run • }

31. NSLS example (2) • proc initData {} { • global A B C D • set A { • { -1 -1 1 1 -1 -1 } • { -1 1 -1 -1 1 -1 } • { -1 1 1 1 1 -1 } • { -1 1 1 1 1 -1 } • { -1 1 -1 -1 1 -1 } • { -1 1 -1 -1 1 -1 } • } • set B { • { 1 1 -1 -1 -1 -1 } • { 1 1 -1 -1 -1 -1 } • { 1 1 1 1 -1 -1 } • { 1 1 -1 -1 1 -1 } • { 1 1 -1 -1 1 -1 } • { 1 1 1 1 -1 -1 } • } • … }

32. NSLS example (3) • proc trainNetwork {} { • global A B C D • memorize $A • memorize $B • memorize $C • memorize $D • } • proc NslMain {} { • global A B C D • puts "Initializing" • initData • puts "Training" • trainNetwork • puts "Testing" • for { set i 10 } { $i<20 } { incr i } { • puts "Testing with distortion $i" • test $A $i • } • nsl set hopfield.dis 0 • } • NslMain

33. Downloading NSL • First, you will need to install the latest Java SDK; get it directly from Sun at http://java.sun.com/j2se/1.3/. • Once this is setup and working, download the entire NSL tree from http://www-scf.usc.edu/~csci564/nsl.tar.gz • Extract the archive (Winzip or Pkzip). • Edit the file "NSL3_0_n\resume.bat" such that it matches your environment (you will have to specify the path where you installed Java, where you installed NSL, etc). • Execute the resume batch file before beginning a NSL session.

34. Downloading NSL (2) • @echo off • echo Initializing NSL environment variables • set NSLJ_ROOT=C:\salvador\NSL3_0_n • set JAVA_HOME=C:\jdk1.3 • set NSL_OS=windows • echo Updating path and classpath • set PATH=%JAVA_HOME%\bin;%NSLJ_ROOT%;%PATH% • set CLASSPATH=%NSLJ_ROOT%;.;%NSLJ_ROOT%\nslj\src\main; • %NSLJ_ROOT%\nslj\src\nsls\jacl; • %NSLJ_ROOT%\nslj\src\nsls\tcljava • @echo on

35. References • A Weitzenfeld, MA Arbib and A Alexander, 2002, NSL Neural Simulation Language, MIT Press (in press) • An old version is at: • http://www-hbp.usc.edu/_Documentation/NSL/Book/TOC.htm • For any NSL related questions and bug reports, please send me an email at smarmol@usc.edu