Greetings from NeuroDimension!
Makers of NeuroSolutions, the Neural Network Simulation Environment.
This edition of the newsletter highlights NeuroSolutions interactivity with other products and introduces you to radial basis function (RBF) networks.
In this issue you’ll find:
What’s New?
Designing Neural Networks
NeuroSolutions Tip Box
Did You Know?
Customer Spotlight
Note: You are receiving this newsletter because you requested to stay informed concerning new developments at NeuroDimension. If you would like to stop receiving these newsletters, please see the bottom of this page for removal instructions.
What’s New?
NeuroSolutions for Excel Patch for Microsoft Excel 2000
We've created a patch to enable NeuroSolutions for Excel to work with Microsoft Excel 2000. If you've recently upgraded to Microsoft Excel 2000, please download the patch from the following location.
http://www.nd.com/support/excel2000_patch.htm
Designing Neural Networks
This Month: Radial Basis Function (RBF) Networks
Most people know about the multi-layer perceptron (MLP) and its broad ranging abilities. The MLP is the workhorse of the neural network field. The radial basis function (RBF) network is similar in its abilities, but approaches the problem differently. An RBF typically has the same goals as an MLP (e.g. classification or function approximation) but uses two distinct "layers" to implement these goals. The first layer is composed of adaptable basis functions, which are usually Gaussians. The location and width of the basis functions are adapted so that they appropriately "cover" the input space. The output of these basis functions becomes your new representation of the input, which is then input to a simple supervised network. The supervised network is then trained to perform the appropriate task.
The obvious question is "Why would you use an RBF instead of an MLP?" A simple way to understand the RBF network is to think that the basis functions are doing a clustering of the input. Inputs that are similar will "fire" similar basis functions, producing very similar inputs to the supervised network. Because you are preprocessing the input to the supervised network, it can typically be much simpler (in fact, often linear) and yet still solve the problem. Since the supervised network is simpler and smaller it trains much faster. And since the RBFs are local in nature (they cluster local areas in the input space), RBFs have a tendency to generalize better.
The next question is "When should I use an RBF instead of an MLP?" The main difficulty with RBFs is the appropriate "covering" of the input space. If you have a large dimensional input (e.g. 10+ columns of data) then you will need a great deal of (multi-dimensional) Gaussians to fill that large input space. As an example, imagine trying to "cover" a 2 dimensional square with small circles. It will take approximately N2 circles to cover the square. Now imagine trying to cover a 3 dimensional cube with small spheres. It will take approximately N3 "balls" to cover the input space. The number of basis functions increases exponentially with the dimensionality of the input. For this reason, MLPs tend to perform better with large dimensional inputs and RBFs perform better on small input spaces. In addition, since the RBFs are essentially clustering the data, RBFs will perform better if the input data can be naturally clustered into regions. When in doubt, try both methods and see which one works better.
To create an RBF network in NeuroSolutions using the NeuralWizard, change the neural model from "Mutlilayer Perceptron" to "RBF/GRNN/PNN Network".
NeuroSolutions Tip Box
This Month: OLE Automation
NeuroSolutions is a fully compliant OLE Automation Server. This means that NeuroSolutions can receive control messages from OLE Automation Controllers, such as Visual Basic (VB), Excel, Access, and Delphi. Writing a fully functioning VB program to control NeuroSolutions is as simple as recording a NeuroSolutions macro, clicking the "Copy as VB" button (from the MacroWizard), and pasting the resulting VB code into the desired VB application.
A typical VB application might simply set a network's parameters, run the network, then retrieve the network's output. However, advanced programmers may want to create applications to run complex batch experiments for network optimization. There is virtually no limit to the level of control you have over NeuroSolutions from your external program.
Two sample OLE projects are included with NeuroSolutions v3.02 -- one for Visual Basic and the other for Visual C++. Please refer to the "OLE Automation" section of the on-line help for an introduction to these projects.
Did You Know?
This Month: Using GeneticServer to Optimize Neural Network Parameters
Neural network performance can often be improved by tweaking some of the neural networks’ key parameters. There are several rules of thumb for setting the initial values for the various neural network parameters. However, rules of thumb are meant to work well in general and do not usually result in the best performance for each individual problem. Finding the best set of neural network parameters can be a very laborious task when attempted manually or a very time consuming task if performed exhaustively through programmatic control of NeuroSolutions.
With the release of GeneticServer, finding the optimum network parameters is now much easier and less time consuming. It is very simple task to write a Visual Basic program that uses the GeneticServer to optimize the network parameters for a breadboard in NeuroSolutions. A sample Visual Basic program is available from our web site that can be used with GeneticServer to optimize the step sizes and the number of hidden layer processing elements for an MLP built with the NeuralWizard. This example code can easily be customized to suit your specific needs. It is available at: http://www.nd.com/public/demo/GeneticExample.zip
Customer Spotlight
This Month: Internet Search
Karl Kurbel, Kirti Singh, Frank Teuteberg
Europe University Viadrina Frankfurt (Oder), Germany
A database of business Internet applications developed at Europe University Viadrina is in the process of being filled with "interesting" WWW applications. As the number of WWW sites is huge and still growing fast, the question is how to find the right applications for the database. In this paper, a neural network approach is proposed to automate the process of searching and selecting candidate applications. 23 configurations of neural networks have been tested: 15 versions of the multi-layer perceptron, four generalized feed-forward networks and four modular networks. Results from training and testing those networks are presented and discussed.
NOTE: This is just an abstract of the application summary. The entire summary is available at:
http://www.nd.com/application%20summaries/appsum-internet.htm
Want to have your solutions spotlighted? We strongly encourage our customers to send their 1-2 page application summaries to submissions@nd.com so that we may post them on our web site at: http://www.nd.com/appliactionsum.htm. In each newsletter, we’ll spotlight a new solution and include a link for people to get more information.
Comments or Suggestions?
We appreciate your feedback! Please send us your comments or suggestions concerning this newsletter, our web site, or part of the NeuroDimension product line. Write to us at: feedback@nd.com
Have questions about NeuroDimension products or training services? Send your questions to: info@nd.com
This issue and previous issues of this newsletter are available on the NeuroDimension web site at: http://www.nd.com/mailinglist.htm
If you would prefer not to receive these newsletters or subsequent product updates from NeuroDimension, please reply to this letter with the subject heading changed to the word REMOVE.
Thank you again for your support of NeuroDimension products!