survey performed by Juha Ikonen, May 21st 1999
SOM_PAK is a collection of tools for implementing the SOM (or
Kohonen Network) algorithm. The program package has been developed to
demonstrate the implementation of the SOM algorithm and to ease first
experiments. It is created by the SOM Programming Team of the Helsinki
University of Technology and can be considered as the original SOM
- The programs are fast but a bit difficult to use.
- Old-fashion or UNIX-style user interface.
- Map parameters are rather limited such as dimensionality, topology.
- Lacks visualization functions
Disclaimer: If any information on this page simply is not
true, please tell us about it and we'll correct it ASAP.
Disclaimer: The opinions and observation herein should be
considered personal of the person having performed by the survey, at
the time of the survey. They do not reflect any official standing of
his employer, of the Laboratory of Computer and Information Science or
the Neural Networks Research Center.
||SOM_PAK Version 3.1
||Program is free for scientific purposes and is available at
SOM Programming Team of the Helsinki University of Technology
Laboratory of Computer and Information Science
Rakentajanaukio 2 C, SF-02150 Espoo
||An interactive monitoring tool using SOM algorithm.
||UNIX and MS-DOS. Program is distributed as source code in C language, pre-compiled
binaries for MS-DOS are also available.
||Command line interface. The package contains 11 separate programs or tools, user must
know which tools to use to achieve desired results. The approach can not be considered
user-friendly, although it may be very effective and convenient in many applications.
||PostScript/text file, which also reveals the basic theory of SOM.
||Standard algorithm. Data vectors can be forced to specified points in resulting map.
[is implementation correct?]
||2-dimensional map grid
minimum size: 1x1
maximum size: 74x74
|Map lattice and shape
||Map lattice: rectangular or hexagonal
Map shape: rectangular
||Function type: Bubble or Gaussian
|Neighborhood size (h):
Parameters: initial value (decreases automatically to one during trainig)
|Learning rate (alpha):
Type: linear or inverse (1/t)
Parameters: initial value (decreases automatically to zero during training)
[Windows NT 4.0, 333 MHz Pentium II, 128 MB RAM]
|for ~3000 13-dim data samples, 4 epochs
training time: 36 seconds
final average quantization error: 0.885901
final topographic error: not applicable
||ASCII, compressed (gzip) ASCII, stdin
|Data handling and selection
||No data handling and selection tools are provided
||ASCII or compressed (gzip) ASCII for data, PostScript for images. Output can be sent
to stdout instead of a regular file.
||Simple: labels can be included in a calibration data set. BMU's for calibration data
samples are sought and labeled with the labels included.
|Inspection of neurons
||Simple: weights of a neuron can be viewed from map file, no visualization tool is
||Clusters: not applicable
Sammon mapping: output to data file and PostScript image.
U-matrix: PostScript image
||Component planes: single plane or all planes.
||Single: BMU can be sought for a data vector, corresponding
label will be attached if defined. Quantization error is also
calculated for the data vector. Projection data is stored in a
file, no visualization tool is provided.
trajectory of BMU's for a set of data vectors can be formed and
saved as a PostScript image.
Monday, 09-Oct-2000 12:53:09 EEST