Kohonen Training Criteria

Kohonen Training Criteria

Top  Previous  Next

_bm41

 

Pattern Selection

You need to click on the button which sets the method you want the training patterns presented to the network.

 

Rotation - This option selects training patterns in order.  Use when like training patterns are dispersed evenly through the training set.

 

Random - This option randomly chooses the training patterns, although it does not guarantee that every pattern will be chosen an equal number of times.  If you select random presentation, the category contents will tend to "bounce around" even when the learning rate is very small.  This is because the random number generator usually does not select all of the patterns during a single epoch, and will select some patterns multiple times.  Consequently, during training the graphics presentation pie chart may be showing multiple copies of some patterns and not other patterns.

 

These are guidelines rather than rules and a user should try different combinations to see which works the best with a particular application.

 

Distance Metric

You need to click on the button which sets the distance metric you want to use during network processing.  Kohonen networks work by clustering patterns based upon their distance from each other.

 

Vanilla or Euclidean distance metric is recommended for most networks because it works the best.   If you use Vanilla, the output of the network is the square of the distance between the pattern and the weight vector for that neuron, therefore the winner is the neuron with the minimum activation.

 

Normalized takes arrays which are linear multiples of one another into the same normalized array (e.g., [1,2,6] and [2,4,12] will both be normalized to [.156, .312, .937]).   You should only use the normalized distance metric when the values for all of the inputs in a pattern are in the same range.  This method must be used with care and is not usually the preferred method.

 

Missing Values

See Missing Values for more details.