DemoGNG-1.5
所属分类:人工智能/神经网络/深度学习
开发工具:Java
文件大小:1220KB
下载次数:21
上传日期:2006-03-28 06:20:31
上 传 者:
sailingyu
说明: 关于自组织神经网络的一种新结构程序,并包含了其它几种神经网络的程序比较
(of self-organizing neural network structure of a new procedure, and contains several other neural network procedures that)
文件列表:
DemoGNG-1.5 (0, 1998-10-19)
DemoGNG-1.5\audio (0, 1996-06-21)
DemoGNG-1.5\audio\computer.au (21745, 1996-06-21)
DemoGNG-1.5\audio\drip.au (759, 1996-06-21)
DemoGNG-1.5\audio\drummer.au (19352, 1996-06-21)
DemoGNG-1.5\Changes (3555, 1997-03-10)
DemoGNG-1.5\CHL.html (2923, 1998-10-19)
DemoGNG-1.5\CHL_0.html (2929, 1998-10-19)
DemoGNG-1.5\CHL_1.html (2926, 1998-10-19)
DemoGNG-1.5\CHL_10.html (2924, 1998-10-19)
DemoGNG-1.5\CHL_11.html (2924, 1998-10-19)
DemoGNG-1.5\CHL_12.html (2932, 1998-10-19)
DemoGNG-1.5\CHL_2.html (2924, 1998-10-19)
DemoGNG-1.5\CHL_3.html (2923, 1998-10-19)
DemoGNG-1.5\CHL_4.html (2933, 1998-10-19)
DemoGNG-1.5\CHL_5.html (2933, 1998-10-19)
DemoGNG-1.5\CHL_6.html (2932, 1998-10-19)
DemoGNG-1.5\CHL_7.html (2928, 1998-10-19)
DemoGNG-1.5\CHL_8.html (2924, 1998-10-19)
DemoGNG-1.5\CHL_9.html (2931, 1998-10-19)
DemoGNG-1.5\ComputeGNG.class (70433, 1998-10-19)
DemoGNG-1.5\ComputeGNG.java (120120, 1998-10-19)
DemoGNG-1.5\COPYING (17982, 1996-06-21)
DemoGNG-1.5\DemoGNG.class (30773, 1998-10-19)
DemoGNG-1.5\DemoGNG.java (63728, 1998-10-19)
DemoGNG-1.5\DemoGNG.zip (62590, 1998-10-19)
DemoGNG-1.5\DemoGNGcode.html (4940, 1998-10-19)
DemoGNG-1.5\discreteData.tcl (675, 1996-09-26)
DemoGNG-1.5\doc (0, 1998-10-19)
DemoGNG-1.5\doc\AllNames.html (21023, 1998-10-19)
DemoGNG-1.5\doc\DemoGNG.html (43076, 1998-10-19)
DemoGNG-1.5\doc\images (0, 1996-06-21)
DemoGNG-1.5\doc\images\blue-ball-small.gif (255, 1996-06-21)
DemoGNG-1.5\doc\images\blue-ball.gif (925, 1996-06-21)
DemoGNG-1.5\doc\images\class-index.gif (1497, 1996-06-21)
DemoGNG-1.5\doc\images\constructor-index.gif (1711, 1996-06-21)
DemoGNG-1.5\doc\images\constructors.gif (1565, 1996-06-21)
DemoGNG-1.5\doc\images\cyan-ball-small.gif (255, 1996-06-21)
DemoGNG-1.5\doc\images\cyan-ball.gif (925, 1996-06-21)
DemoGNG-1.5\doc\images\error-index.gif (1438, 1996-06-21)
... ...
-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Sorry for the bad quality of this README file. It was generated from the
dvi-file of the documentation. Please have a look at the corresponding
html- or PostScript file. Thank you, Hartmut Loos.
-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
DemoGNG v1.5
Hartmut S. Loos 1
and
Bernd Fritzke y2
1Systems Biophysics
Institute for Neural Computation
Ruhr-Universit"at Bochum
2Neural Computation Group
Artifical Intelligence Institute
Computer Science Department
Technical University Dresden
October 19, 19***
Abstract
DemoGNG, a Java applet, implements several methods related to com-
petitive learning. It is possible to experiment with the methods using various
data distributions and observe the learning process. A common terminology
is used to make it easy to compare one method to the other. Thanks to the
Java programming language the implementations run on a very large number
of platforms without the need of compilation or local adaptation. Hopefully,
the experimentation with the models will increase the intuitive understanding
and make it easier to judge their particular strengths and weaknesses.
_____________________________________
http://www.neuroinformatik.ruhr-uni-bochum.de/ini/PEOPLE/loos
yhttp://pikas.inf.tu-dresden.de/ fritzke
1
Contents
1 Introduction 3
2 System Requirements 3
3 Installation 3
4 Starting the Java Applet 4
5 Manual 4
5.1 Network Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
5.2 Drawing Area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
5.3 General Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
5.3.1 Buttons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
5.3.2 Checkboxes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
5.3.3 Pull-Down Menus . . . . . . . . . . . . . . . . . . . . . . . . 8
5.4 Model Specific Options . . . . . . . . . . . . . . . . . . . . . . . . . . 9
5.4.1 LBG, LBG-U . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
5.4.2 Hard Competitive Learning . . . . . . . . . . . . . . . . . . . 9
5.4.3 Neural Gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
5.4.4 Competitive Hebbian Learning . . . . . . . . . . . . . . . . . 11
5.4.5 Neural Gas with Competitive Hebbian Learning . . . . . . . 11
5.4.6 Growing Neural Gas, Growing Neural Gas with Utility . . . . 11
5.4.7 Self-Organizing Map . . . . . . . . . . . . . . . . . . . . . . . 12
5.4.8 Growing Grid . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
6 Wishlist 13
References 14
7 Change log 15
List of Figures
1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 Pull-down menu with all available models . . . . . . . . . . . . . . . 5
3 Example screenshots of the available models . . . . . . . . . . . . . . 6
4 Drawing Area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
5 Options for all models . . . . . . . . . . . . . . . . . . . . . . . . . . 7
6 Some examples for activated checkboxes. . . . . . . . . . . . . . . . . 8
7 General pull-down menus . . . . . . . . . . . . . . . . . . . . . . . . 8
8 Overview of the available probability distributions . . . . . . . . . . 10
2
1 Introduction
In the area of competitive learning a rather large number of models exist which have
similar goals but differ considerably in the way they work. A common goal of those
algorithms is to distribute a certain number of vectors in a possibly high-dimensional
space. The distribution of these vectors should reflect (in one of several possible
ways) the distribution of input signals which in general is not given explicitly but
only through sample vectors.
2 System Requirements
To run DemoGNG, a Web browser with Java enhancements (e.g. Netscape 2.0 or
newer, Microsofts Internet Explorer, ...) is needed or, e.g. the Sun Microsystems
appletviewer from the Java Developers Kit (JDK).
3 Installation
Download the compressed tar version1, save it to a file, say DemoGNG-1.5.tar.gz,
and then extract the files with
% gunzip DemoGNG-1.5.tar.gz
% tar xvf DemoGNG-1.5.tar
Instead, you can download DemoGNG in zip-format2.
After uncompressing and unpacking you should have the following:
_____________________________________1
ftp://ftp.neuroinformatik.ruhr-uni-bochum.de/pub/software/NN/DemoGNG/DemoGNG-
1.5.tar.gz2
ftp://ftp.neuroinformatik.ruhr-uni-bochum.de/pub/software/NN/DemoGNG/DemoGNG-
1.5.zip
3
COPYING copy of the GNU General Public License
Changes change history for DemoGNG
ComputeGNG.java Java source for class ComputeGNG
DemoGNG.java Java source for class DemoGNG
EdgeGNG.java Java source for class EdgeGNG
EdgeVoronoi.java Java source for class EdgeVoronoi
FPoint.java Java source for class FPoint
GraphGNG.java Java source for class GraphGNG
GridNodeGNG.java Java source for class GridNodeGNG
HalfEdgeVoronoi.java Java source for class HalfEdgeVoronoi
LineGNG.java Java source for class LineGNG
ListElem.java Java source for class ListElem
ListGNG.java Java source for class ListGNG
NodeGNG.java Java source for class NodeGNG
SelGraphics.java Java source for class SelGraphics
SiteVoronoi.java Java source for class SiteVoronoi
DemoGNGcode.html starting page for source documentation
GenerateHTML.tcl a tcl-script to generate some HTML-pages with different
parameters
Makefile Makefile for Unix make
README the ASCII-version of this document
SwitchGNG.html the main HTML-page
audio/ subdirectory containing audio files
doc/ subdirectory containing the javadoc generated documenta-
tion of DemoGNG with the needed images
tex/ subdirectory containing the LaTeX source of the manual
with the DVI-, PS- and HTML-format
Type make world to generate some necessary files (all these files are included
in the archive).
*.html needed for SwitchGNG.html
*.class compiled classes
doc/*.html javadoc generated HTML-files
4 Starting the Java Applet
Start the Web browser with the main HTML-page.
Example:
% netscape SwitchGNG.html
Important: Make sure you have Java support enabled.
Instead, you can also use an appletviewer with one of the generated HTML-files.
Example:
% appletviewer GNG.html
Important: The appletviewer requires an applet included in the Web page, so you
must choose one of the generated HTML-files.
5 Manual
After the Java applet has been launched, the DemoGNG main window appears.
This window can be divided into four regions:
1. Network Model
4
Figure 1: Overview
2. Drawing Area
3. General Options
4. Model Specific Options
5.1 Network Model
Figure 2: Pull-down menu with all available models
The first part shows the actual algorithm. To select another algorithm or restart
the current, click on the pull-down menu and choose the desired algorithm. The
following algorithms are available:
fflLBG (Linde et al., 1***0)
fflLBG-U (Fritzke, 1997b)
fflHard Competitive Learning (standard algorithm)
- with constant learning rate
- with decreasing learning rate
fflNeural Gas (Martinetz and Schulten, 1991)
fflCompetitive Hebbian Learning (Martinetz and Schulten, 1991; Martinetz,
1993)
5
fflNeural Gas with Competitive Hebbian Learning (Martinetz and Schulten,
1991; Martinetz and Schulten, 1994)
fflGrowing Neural Gas (Fritzke, 1994; Fritzke, 1995a)
fflGrowing Neural Gas with Utility (Fritzke, 1997a)
fflSelf-Organizing Map (Kohonen, 1***2)
fflGrowing Grid (Fritzke, 1995b)
LBG, LBG-U Hard Competitive Learning
Neural Gas Competitive Hebbian Learning
Neural Gas with CHL Growing Neural Gas, GNG-U
Self-Organizing Map Growing Grid
Figure 3: Example screenshots of the available models
5.2 Drawing Area
The drawing area shows the network for the selected algorithm. The geometric
figure in the background of the area reflects the probability distribution (see Section
6
Figure 4: Drawing Area
5.3.3 for more information). In every phase of the algorithm you can select a node
and drag it to an arbitrary location within the drawing area. Additional information
is displayed in the corners of this region:
upper left Number of input signals occured so far
upper right Version number
lower left Number of nodes
lower right Some additional information
5.3 General Options
Figure 5: Options for all models
This region contains the non-specific parameters for all algorithms. There are three
kinds of interface elements: buttons, checkboxes and pull-down-menus.
5.3.1 Buttons
Start Starts/Continues a stopped run.
Stop Stops a calculation to modify parameters or move nodes (these
modifications can also be done while the simulator runs).
Reset Resets the algorithm, but leaves all parameters unchanged (to
restart an algorithm with default parameters see Section 5.1 for
more information).
7
Teach mode Voronoi diagram Error graph
Figure 6: Some examples for activated checkboxes.
5.3.2 Checkboxes
Teach Toggles teach mode. In the teach mode the algorithm is slowed
down and more information is displayed depending on the algo-
rithm (default: off).
Signals Toggles display of signals. The most recently generated signals
are shown (default: off).
Voronoi Toggles display of Voronoi diagram (default: off).
Delaunay Toggles display of Delaunay triangulation (default: off).
Error Graph Toggles the error graph. It is displayed in a separate window and
shows the mean square error of the selected model (default: off).
Nodes Toggles display of nodes (default: on).
Edges Toggles display of edges (default: on).
Random Init If this is switched on, initial node positions can lie outside the
region of positive probability density (default: off).
White Switches the background of the drawing area to white. This is
useful for making hardcopies of the screen (default: off).
Sound Toggles sound (default: off).
5.3.3 Pull-Down Menus
Probability Distribution (max.) Nodes Display Speed
Figure 7: General pull-down menus
8
prob. Distrib. Selects one of the available probability distributions. The choosen
distribution is displayed in the drawing area (see Section 5.2 for
more information). The following distributions are provided: The
convex uniform ones Rectangle, Circle and Ring. The clustered
uniform ones UNI, Small Spirals, Large Spirals and UNIT.
The non-uniform HiLo Density which consists of a small and a
large rectangle. Each of these rectangles gets 50% of the signals.
The discrete distribution Discrete, which consists of 500 data
vectors generated from a number of Gaussian Kernels.
And the non-stationary uniform distributions Move & Jump,
Move, Jump and Right MouseB. The first three distributions
are moving automatically, for the last one click on the right mouse
button and place the distribution where you want. Or hold the
right mouse button and the distribution will follow the mouse
pointer.
(max.) Nodes Selects the number of nodes (Growing Neural Gas and Grow-
ing Grid: maximum number of nodes).
Display Selects the update interval for the display.
Speed Selects an individual speed depending on the machine and/or
browser.
Select a slow speed for good interaction with the program and
slower program execution.
Select a fast speed for slow interaction and fast program execution.
The most suitable setting depends on your local hard- and soft-
ware.
5.4 Model Specific Options
This region shows the model specific parameters. Each time a new model is selected,
the necessary parameters are displayed. For a complete description of the models
and their parameters look at the technical report Some Competitive Learning Meth-
ods by Bernd Fritzke which is available in HTML3- and in Postscript-format4.
5.4.1 LBG, LBG-U
_________________________
_ Number of Signals _The number of input signals (only discrete distributions).
_________________________ _
___________
__LBG-U____ _Switches from LBG to LBG-U and back.
5.4.2 Hard Competitive Learning
_____________
__Variable____Switches from a constant to a variable learning rate.
___________
_ epsilon _This value (ffl) determines the extent to which the winner is adapted
____________
towards the input signal (constant learning rate).
_____________
_ epsilon_i _epsilon initial (ffl ).
______________ i
_____________
_ epsilon_f _epsilon final (ffl ).
_____________ _ f
__________
__t_max____The simulation ends, if the number of input signals exceeds this value
(tmax ).
_____________________________________3
4http://www.neuroinformatik.ruhr-uni-bochum.de/ini/VDM/research/gsn/JavaPaper
ftp://ftp.neuroinformatik.ruhr-uni-bochum.de/pub/software/NN/DemoGNG/sclm.ps.gz
9
UNI Small Spirals UNIT
Rectangle Large Spirals HiLo Density
Circle Ring Discrete
Move & Jump Move Jump
Figure 8: Overview of the available probability distributions
The variable learning rate is determined according to
ffl(t) = ffli(fflf =ffli)t=tmax:
5.4.3 Neural Gas
_____________
__lambda_i___ _lambda initial (i).
______________
__lambda_f_____lambda final (f ).
_____________
_ epsilon_i _epsilon initial (ffl ).
______________ i
_____________
_ epsilon_f _epsilon final (ffl ).
_____________ _ f
__________
__t_max____The simulation ends, if the number of input signals exceeds this value
(tmax ).
The reference vectors are adjusted according to
w i = ffl(t) h (ki( ; A)) ( w i)
10
with the following time-dependencies:
(t) = i(f =i)t=tmax
ffl(t) = ffli(fflf =ffli)t=tmax
h (k) = exp (k=(t)):
5.4.4 Competitive Hebbian Learning
This implementation requires no model specific parameters. In general one would
have a maximum number of time steps (tmax ).
5.4.5 Neural Gas with Competitive Hebbian Learning
_____________
__lambda_i___ _lambda initial (i).
______________
__lambda_f_____lambda final (f ).
_____________
_ epsilon_i _epsilon initial (ffl ).
______________ i
_____________
_ epsilon_f _epsilon final (ffl ).
_____________ _ f
__________
__t_max____The simulation ends, if the number of input signals exceed this value
(tmax ).
__________
_ edge_i _Initial value for time-dependend edge aging (T ).
___________ i
__________
_ edge_f _Final value for time-dependend edge aging (T ).
__________ _ f
Edges are removed with an age larger than the maximal age T (t) whereby
T (t) = Ti(Tf =Ti)t=tmax:
The reference vectors are adjusted according to
w i = ffl(t) h (ki( ; A)) ( w i)
with the following time-dependencies:
(t) = i(f =i)t=tmax
ffl(t) = ffli(fflf =ffli)t=tmax
h (k) = exp (k=(t)):
5.4.6 Growing Neural Gas, Growing Neural Gas with Utility
________ ... ...
近期下载者:
相关文件:
收藏者: