Greetings from NeuroDimension!
Makers of NeuroSolutions, the Neural Network Simulation Environment.
This issue of the newsletter highlights our genetic products and data classification techniques. As usual, it also includes other tips and techniques that will help you to get the most out of NeuroSolutions from the professionals at ND.
In this issue you’ll find:
What’s News?
* Early Course Registration Discount Ends Soon
* Genetic Server and Library - Information Updated
Designing Neural Networks
* Unsupervised Neural Networks for Preprocessing
NeuroSolutions Tip Box
* Configuring Breadboards with Files of Alternate Data Types
Did You Know?
* Creating User-Defined Genetic Operators With GeneticLibrary
Customer Spotlight
* The Classification of Post Office Outlets According to Risk of Incident
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 newsletter for removal instructions.
What’s News?
Early Course Registration Discount Ends Soon
The next NeuroDimension course offering will be November 8-12, 1999 in Orlando, Florida. Register before October 1st and receive a 10% discount. We continue to receive rave reviews from our attendees, in fact everyone has said that they would recommend our courses to friends. We average a score of 4.6 of 5 when attendees are asked to rate the overall quality of the course. Our overlapping course format allows both novice and advanced users to find a suitable course. Offered courses include: "Introduction to Neural Networks and NeuroSolutions", "Advanced Neural Networks and NeuroSolutions", and "NeuroSolutions for Developers".
For details on the November offerings, or to sign-up immediately from the internet, see http://www.nd.com/course/november.htm
For general ND course information, see http://www.nd.com/course
Genetic Server and Library - Information Updated
What is a genetic algorithm and what can it do for you? If you are interested in finding out more about Genetic Server 1.0 and Genetic Library 1.0, our new products which provide a general purpose API for custom genetic applications, please visit our web site. We've recently updated the site to include more information on features, applications, and genetic algorithm definitions. See http://www.nd.com/products/genetic.htm
Designing Neural Networks
This Month: Unsupervised Neural Networks for Preprocessing
There are two fundamental types of neural networks, supervised and unsupervised. Supervised networks include multi-layer perceptrons, radial basis functions, and all other architectures that are trained with both the input and a desired signal. Unsupervised networks do not use a desired signal, so they extract all of their information from the input via an unsupervised learning rule. Two of these learning rules that are useful for preprocessing are the competitive learning rule and Sanger’s learning rule.
Competitive neural networks segment the input into clusters. These networks are made up of a competitive synapse from the unsupervised palette in NeuroSolutions (typically the conscience full synapse) and an axon (typically a winner-take-all axon). The number of PEs in the axon determines the number of clusters the network will create. Competitive clustering networks can be used as preprocessors for supervised networks that will do classification. The advantage is that the clustering significantly simplifies the input to the supervised network. You must be careful, however, since if two different classes of inputs are clustered together, the supervised network will not be able to separate them.
Sanger’s rule implements a principal component analysis (PCA) network. It consists of a Sanger’s synapse and an axon (typically a plain, vanilla axon). Sanger’s network compresses the input data down to its N largest "components", where N is defined by the number of PEs in the axon. A PCA network does the best linear compression of data from the input space to an N-dimensional output space. This network is useful when you have input with a very large dimensionality (e.g. many columns), but the information may be highly correlated between the columns. Thus, the actual information contained in the input can be efficiently compressed. As in the competitive network, the PCA network can be used as a preprocessor for a supervised network. The PCA network can significantly reduce the dimensionality of the input to the supervised network, creating a smaller, easier to train network.
The PCA hybrid (PCA followed by a supervised) network can be created automatically in the neural wizard by selecting the Principal Component Analysis. The competitive hybrid network can be built by selecting the RBF network and replacing the Gaussian axon (the cracked one) with a winner-take-all axon or a softmax axon.
NeuroSolutions Tip Box
This Month: Configuring Breadboards with Files of Alternate Data Types
NeuroSolutions supports several different data types as sources for its data. However, the NeuralWizard currently requires that you use column-formatted ASCII text files for the input and desired data of the neural network. The other data types that NeuroSolutions currently supports are the following:
Plain ASCII – Text files that do not have column headings. The data values are normally delimited by spaces, tabs or commas and it is not required that they be aligned into columns.
Bitmap – Either 16 or 256 color bitmap (*.bmp) images. Each pixel may be translated into either one grayscale value or three RGB values (0 to 255).
Binary – Data values are stored as 4-byte floating point numbers. This is the same file format that NeuroSolutions uses to read in the data during a simulation -- also referred to as a pattern file (*.nsp).
Setting up a neural network to use one of these file types is very easy. Here are the steps you need to take:
When using column-formatted ASCII files, the number of data channels is determined based on the selected columns. With these alternate file types, the number of data channels must be specified by changing the number of rows (and columns if using bitmaps) of the Axon or Criterion that the File is attached to.
Did You Know?
This Month: Creating User-Defined Genetic Operators With GeneticLibrary
GeneticLibrary allows the user to create and utilize custom genetic operators. Specifically, this feature is available for the selection, crossover, and mutation operators. This article will demonstrate this functionality by creating a custom selection operator. A selection operator is used to select which chromosomes will be used to create the next generation.
Assume that you would like to develop and use a new selection operator named TopTenPercentSelection that randomly chooses a chromosome from the top ten percent (according to fitness) of all of the chromosomes in the current population. This can be accomplished in three easy steps.
First, add the following prototype to the top of your source file:
Chromosome & TopTenPercentSelection(Population &population);
From this prototype, you can see that a custom selection operator is passed the current population and is expected to return the selected chromosome.
Next, add the following function to your source file:
Chromosome & TopTenPercentSelection(Population &population)
{
// Sort chromosomes from best to worst according to fitness
population.sort();
// Randomly choose a chromosome from the top 10% of the population
int lastIndex = (int) (0.10f * (population.getNumberOfChromosomes() - 1));
return population.chromosomes(ga.randomInt(0, lastIndex));
}
This function sorts the population (best fitness to worst fitness) then randomly selects and returns a chromosome from the top ten percent of this sorted population.
Finally, to use this custom selection operator, add the following code to your function that implements the genetic algorithm:
ga.setSelectionType(USER_DEFINED_SELECTION, &TopTenPercentSelection);
This call sets the selection type to user-defined and passes the genetic algorithm the address of your user-defined function. Note: This code assumes that a genetic algorithm object named ga was previously created.
As mentioned previously, customer operators can also be defined for crossover and mutation. See the GeneticLibrary help for more information.
Customer Spotlight
This Month: The Classification of Post Office Outlets According to Risk of Incident
Dr. Dragan Savic, University of Exeter, Exeter, Devon, England
The task consisted of the use of data mining tools of all types to find patterns in a database supplied to us by the Royal Mail. The aim of the project was to evaluate novel and already existing ‘data mining’ tools. Data mining is the extraction of useful information from a database using artificial intelligence algorithms and neural networks were considered as a major part of this process. In many sections of the literature it is commented that neural networks often outperform their counterparts on predictive accuracy and achieving this was considered paramount in this project. The database consisted of a set of attributes about each Post Office outlet and an indication whether that office had been subject to an incident in the past three years. At the time of writing it has not been made clear to the investigator what constitutes such an incident. The task of each of the data mining tools was to discover some connection between the number of incidents and the attributes that are present in the database. If satisfactory accuracy can be achieved, then the classification algorithm (or a group of algorithms) should be able to be interrogated as to the classification of new Post Office outlets to ascertain whether they are at risk from an incident.
NOTE: This is just an abstract of the application summary. The entire summary is available at:
http://www.nd.com/application%20summaries/appsum-risk.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/applicationsum.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!