Why is the size of the input weight matrix sometimes smaller than the input length when training a neural network?

I have a question regarding the size of the inut weight matrix for a neural network. My IW Matrix is smaller than expected and I don't know why. What I do:

net=patternnet(1);
[net,tr]=train(net,inputs,targets);
net.IW              %size of the input weight matrices

ans = 
    [1x14 double]
    []

net.inputs.size     %size of my inputs

ans = 
    [15]

net.layers.size     %size of my hidden and output layer

ans = 
    [1]
    [2]

As far as I understood, the size of my input weight matrix should be 1 (size of hidden layer) by 15 (length of input vectors). I tried it several times with different input sizes, but the size of IW sometimes is equal or 1-2 smaller than my input size.

I want to know why this happens and how I can match the weights to the input variables. Thanks in advance, Antje

[Sean de Wolski]

Are the inputs unique?

[Greg Heath]

What are

size(inputs)

size(targets)

Greg

Answer 1

[Antje]

Oh, finally I got it!

The problem was, that there were some redundant columns in the inputs. I have not seen them, because I have a huge amount of data.

It seems that the training process ignores these columns, but I could not see which of them.

If I get rid of them before the training, than there are no deviations in size of the input and size of the weight matrix.

[Greg Heath]

Redundant columns won't produce that behavior.

Do you mean constant rows?

[Antje]

Oh yes, there are some constant rows. I forgot, that I transposed them for better visibility. Sorry.

[Steven Lord]

For more information about the pre- and post-processing functions applied to network inputs and outputs, take a look at this page in the documentation and the documentation page for the removeconstantrows function described on that page.

If you're using an older release of Neural Network Toolbox the specific function being applied may have a different name, but I believe the same general idea applies.

[enjy fikry]

how can i stop that from happening ? i don't want the training process to ignore these constant columns

[Greg Heath]

You should.

They have zero variance.

Therefore they cannot contribute to learning.

However, they can confuse those who do not understand this.

Hope this helps.

Greg