View this message in a browser.

Neurosolutions Newsletter
Visit neurosolutions.com | Download Evaluation Volume 7, No.1
This is a Neurosolutions Newsletter, which you are receiving because you requested to stay informed about new developments at NeuroDimension. You can also access current and past issues of Neurosolutions Newsletters online.
What to after Building a SOM (clustering) Network

The NeuralExpert makes it quite easy to build a Kohonen Self-Organizing Map (SOM) network to partition a set of input data into a number of groups (clusters). After training a SOM, each cluster should represent sets of input rows that have similar characteristics. Once you've built and trained the network, it may not necessarily be intuitive what to do next to interpret the results.

The first thing we recommend is to read the previous newsletter article on SOMs:

http://www.nd.com/newsletter/july_2003.htm

A sample dataset can be downloaded here.

Next, try working through the following list of steps with the SOM breadboard you built:

  1. View the Histogram using an ImageViewer probe in addition to the Hinton probe
    1. Stamp a MatrixViewer on top of the Hinton icon
    2. Open the inspector for the ImageViewer, switch to the "Probe" tab and check the "Automatic" switch.
    3. Change the "Display Every" to the number of rows in your data file. If you are unsure of this number, then check this setting for the Hinton probe.
    4. Double-click the ImageViewer icon to display the probe window.
  2. Change the "Epochs/Run" of the StaicControl inspector to a large number like 10000.
  3. Click the Start button to begin training the network.
    1. Monitor the Quantization Metric probe to make sure the number is continually decreasing. This Quantization Error is a measure of the average distance between each data point and the point represented by its "winning" processing element of the map.
    2. If the error rises immediately or jumps around then the learning rate is probably set too high. In this case you should lower the "Step Size" value within the inspector the SquareKohonenFull component.
    3. Stop the training once the Quantization Error flattens out.
  4. Determine where the clusters are on the map
    1. Observe the ImageViewer histogram and determine the light and dark areas of the map. The brightest points are those processing elements that "won" the most competitions and they usually represent the cluster centers. The black areas are processing elements that did not "win" any of the competitions, so these usually represent the boundaries between neighboring clusters.
    2. If the number of clusters is to high or low for your particular application, then you may want to try adjusting the size of the map and re-train the network. Simply change the number of Rows and Cols of the WinnerTakeAllAxon, click Reset and then Start.
  5. Determine which data points belong to which clusters.
    1. If one of your skipped input columns is a record identifier, then use it for annotation:
      1. From the File inspector click the "Customize" button
      2. Selected the heading that represents the record identifier, click the "Annotate" button and click "Close".
      3. Open the inspector for the DataWriter component stamped on the right side of the WinnerTakeAllAxon. From the "DataWriter" tab check the "Display Annotation" checkbox.
    2. Double click on the rightmost DataWriter icon to display the "Winning PE of outputAxon" probe.
    3. Zero the counter (click the "Zero Count" toolbar button).
    4. Step an epoch (click the "Step Epoch" toolbar button). Note: if this button is not visible, select "Customize" from the Tools menu and check the "Control" checkbox.
    5. Observe the value of the "Winner" for a particular row. This the index of the "winning" processing element for that row of input data.
    6. Find the location of that processing element on the map.
    7. Observe the ImageViewer histogram and count the number of pixels from the upper-left corner, reading left to right. For example, if you have a 10x10 map and you are looking for PE #53, the corresponding pixel would be in the 4th column and the 6th row.
    8. Determine which cluster that processing element belongs to. Find other rows that have a "winner" as the same processing element, or a nearby processing element on the map which belongs to the same cluster. These input rows should all have similar characteristics.
If you would prefer not to receive subsequent product updates or newsletters from NeuroDimension, please feel free to advise us by email and we will be happy to update your details on your behalf.