Hi,
I'm using the R nnet package to train a classifier to recognise items which
belong to a particular class and those which don't belong to the class.
I'm supplying nnet with a matrix x containing training examples (in each
row) and a matrix y of targets. The training set is made up of 200 positive
examples and 1000 negative examples. I want to train the network on the same
number of positive and negative examples.
I have a few questions regarding the way this nnet method works...
1. How are input examples selected and shown to the neural network during
training? Are they shown in the order in which they appear in the examples
matrix (x)? Or are they selected randomly? If they are selected randomly, does
the 'weights' parameter influence how regularly items are selected?
For example, if I weighted each of my positive examples with weight 5 and my
negative examples with 1, would the positive examples then be selected and shown
to the network as frequently as the negative examples (since there are 200
compared with 1000)? This is my aim.
2. The MaxNWts parameter seems to suggest that the number of nodes/weights
changes as the network is trained. Is this the case?
I haven't supplied any examples since my questions are of a more theoretical
nature.
Thank you in advance for your comments,
Hannah
[[alternative HTML version deleted]]
