minimum and maximum input value with newff

I am using the neural network toolbox in Matlab to create a feed-forward backpropagation network. I am using the function:

net = newff(input, target, [35], {'tansig', 'tansig'});

In earlier versions of this function, the first parameter was a matrix with minimum and maximum values for every feature, but in the latest version it has been replaced by input and target samples.

I don't know if the minimum and maximum values are exactly the minimum and maximum values of the samples.

For example: if I have this input:

input = [1 5; 2 7]
input=
        1 5
        2 7

Where the columns are the different samples and the rows are the different features. If after training the network I use it with the next input:

input = [2; 8]
input=
        2
        8

8 exceeds the maximum value in the sample (7). Will I have any problem?

[Greg Heath]

Insufficent information:

Regression or classification?

What physical quantities are the inputs and targets?

 [ I N ] = size(input)% [ ? ? ]
 [ O N ] = size(target)%[ ? ? ]
 H = 35      % ?? Seems rather large
 Nw = (I+1)*H +(H+1)*O  % Number of UNKNOWN weights
 Ntrn ~ 0.7*N       % Default number of training samples
 Ntrneq = 0.7*N*O   % Resulting No. of training EQUATIONS

With H=35 are there more unknowns than equations?

[Greg Heath]

Don't worry about it. The basic statistical assumption is that the trn,val and tst subsets can be assumed to be randomly picked from the same probability distribution.

After designing and testing you can investigate the source of any errors.

At this point be more concerned with just using as few hidden nodes as possible that will satisfy your training goal.

Hope this helps.

Greg

[Ramon Domenech]

OK, thanks