Downloads
Documentation
Theory
Related stuff
_________________
CIS Home
Neural Connection 2.0
survey performed by Juha Vesanto, May 29th 1998
Neural Connection 2.0 is a neural networks tool for data analysis, prediction, classification, clustering and various other purposes. It is part of the SPSS program family. It has implementations of Multi-Layer Perceptron, Radial Basis Function, Bayesian Network, Kohonen Network, Closest Class Means Classifier, Regression, Principal Component Analysis, and various data handing tools.
The SOM (or Kohonen Network) is only one of the many tools in the program - and it suffers from it. Although the technical implementation of the SOM algorithm seems correct, the special features of the SOM have not been taken into account very well. Also the control of the training procedure is somewhat dubious. Major deficiencies include:
- map is always square in shape (except when it is 1-dimensional)
- although the neighborhood size can be made to decrease during training, this is not the default
- no map measures are implemented
- map visualization and post processing in general was really very poor
- training was a bit slow
In addition we had major stability problems with the program (in Windows NT 4.0). On the other hand, the GUI made training the maps easy and the presence of several other tools was a big bonus.
Disclaimer: we only experimented with the program for a short time. Therefore some deficiencies (especially regarding postprocessing) may be due to our lack of understanding the program. Still, in a good program such shouldn't happen.
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.
General |
|
| Program name | Neural Connection 2.0 |
|---|---|
| Availability | Commercial product, single license about $1500 Contact information on the web page: http://www.spss.com/software/Neuro/ |
| Purpose | Data analysis and decision support using neural networks |
| Operating system | Windows |
| User interface | GUI, script language Good (in general) Mediocre (regarding the SOM) |
| Documentation | Online help Good (user interface) Mediocre (technical/scientific details) |
SOM features |
|
| map parameters | |
|---|---|
| Teaching algorithm | standard SOM algorithm - possibility of increasing map size during training by 'doubling' the nodes every N training steps - implementation seems ok, but couldn't be properly verified |
| Map size | 1- or 2-dimensional map grid - initially 10x10 maximum size, but by using doubling this can be increased - cannot have more map units than data vectors |
| Map lattice and shape | rectangular lattice, sheet (rectangular) shape |
| Neighborhood function | Function type: bubble (probably) |
| Neighborhood size: Type: fixed / decreasing by a certain percentage each training cycle (possibly: d(t)=d0*(1-p/100)^t) Parameters: initial neighborhood, decreasing rate |
|
| Learning rate: [type (linear, 1/t, other), parameters] Type: decresing (probably in the same way as for neighborhood size) Parameters: inital learning rate, decreasing rate |
|
| Initialization | data sample / random / small random / grid / small grid |
| Distance function | Euclidian / dot product |
| Unknown components | unclear whether these are allowed or not |
| Teaching length | explicit in epochs training can be intercepted |
| efficiency | |
| Speed [Windows NT 4.0, 200 MHz Pentium Pro, 64 MB RAM] | for ~3000 13-dim data samples, 10 epochs training time: 4 minutes (map converged after 2nd epoch, though) |
| Results | seemed ok final average quantization error 1.29 final topographic error 13.5% |
| Efficiency comparison with SOM Toolbox for Matlab. Same computer, same data set, same training length. Training time 300s (sequential training), 30s (batch training). Final quantization error (sequential training) 0.68, final topographic error 1.1%. | |
Usability |
|
| preprocessing | |
|---|---|
| Input formats | Very good assortment: ascii (various kinds of formats), SPSS 6.0/7.0, MS Excel 5.0, Systat 5.0 |
| Data handling and selection | The environment offers very good tools for data
handling. Specifically for SOM, just before training the following
preprocessing operations were provided: none / normalization of each component to unit variance & zero mean / shphering of each vector to unit length |
| postprocessing | |
| Output formats | Very good assortment: ascii (various kinds of formats), SPSS 6.0/7.0 |
| Map measures | none |
| Labeling | no |
| Clustering | by visualization |
| visualization | |
| Inspection of neurons | simple (inspection of weight vectors) |
| Clusters/map shape | some kind of distance matrix (possibly mean distance to the neighbors of the unit) |
| Correlations | component planes, only a single component shown at a time |
| Data projections | no (it's possible to find BMU for vectors in test set, but no visualization tool for this is provided) |
| Markers | no |
| The usability of the SOM was not very good. Training and preprocessing were ok, but postprocessing and map visualization was hard to do. We also ran into severe stability problems, for example when trying to link the Kohonen Network tool to Data output tools. | |
Other notes
- In case of multiple classes, one map per class could be trained.
- Data limits: 750 fields, 15000 records (32000 when "running" applications)
http://www.cis.hut.fi/projects/somtoolbox/links/neural_connection.shtml
somtlbx@mail.cis.hut.fi
Monday, 09-Oct-2000 12:53:09 EEST