How to insert 2D-matrix to a backpropagation neural network?

I am working on speech restoration, I used MFCC to extract the features. now I have 12*57 input matrix and 12*35 target matrix for each audio clip. My question is how can I insert this 2D-matrix to a backpropagation neural network?

Answer 1

[Greg Heath]

N I-dimensional "I"nput column vectors

 [I N ] = size(input)

N O-dimensional "O"utput target vectors

 [O N ]  = size(target)

 help fitnet     % Regression and Curvefitting
 help patternnet % Classification and Target Regression

%H = number of hidden nodes in I-H-O node topology

 net = fitnet(input,target,H);

% You have a problem with I = 57, O = 35, N = 12 because even if you used all 12 images for training you would only have

Ntrn   = N    % 12   input/target pairs
Ntrneq = N*O  % 420   training equations

whereas there are

 Nw = (I+1)*H+(H+1)*O

unknown weights. Therefore, not to have fewer equations than unknowns (Ntrneq >= Nw), the number of hidden nodes H cannot be larger than the upper bound

 Hub = (Ntrneq-O)/(I+O+1) = 4.1

It is more likely that you need more than H=4 hidden nodes. It is obvious from the above equation that decreasing I and/or increasing N is desired.

Alternatives are

 1. Use MSEREG as the training goal
 2. Use TRAINBR as the training function.

Personally, I would try input-variable reduction and/or obtain more input/target pairs.

The best inputs to keep are usually indicated via the linear coefficient model STEPWISEFIT.

Hope this helps.

Thank you for formally accepting my answer

Greg

[nada fady]

Thank you very much for your clear answer. I think I wrote some thing wrong in my question, that is: I have 12*35 matrix input and 12*57 target output, so can input matrix size be smaller than target output size?