The "N"umber of individual "I"-dimensional "I"nputs and corresponding "O"-dimensional "O"utput "t"argets are given by
x = input; size(x) = [ I N ]
t = target; size(t) = [ O N]
I have hundreds of posts in both the MATLAB NEWSGROUP (comp.soft-sys.matlab) and ANSWERS which estimate the maximum number of "H"idden nodes, H <= Hmax to avoid overtraining an overfit net by simply not overfitting.
Since only one hidden layer is sufficient, a straightforward search over random initial weights easily determines the minimum number of hidden nodes for a given performance. I use
mean-square-error(training set output) <= 0.01*mean( training set target variance)
(with the obvious assumption that the validation and test subsets have statistics similar to that of the training subset). In other words, I try to do 100 times better than assuming the output is a constant equal to the mean of the target.
Adding the search word "greg" tends to yield my most appropriate posts in most neural network topics.
Hope this helps
Thank you for formally accepting my answer
Greg
