Practical Neural Network Recipies in C++

340 Chapter 18

training case least represented in the network. It is reasonable to

assume that this training pattern deserves a neuron of its very own,

rather than being forced to share a neuron with patterns dissimilar to

it.

The second step is to present the network with the outlying

case found in the first step. Examine the activation level of all

neurons that did

not win

for any training

case.

Select the neuron from

among these that has maximum activation. This is the neuron that

we will choose to represent that case.

The final step is to use that case to set the weight vector for

that neuron. This is trivial, in that we simply copy th ...