Hi Ali,

I suggest that you whiten your data while performing the data dimensionality reduction. Matlab source code is available by checking for:

• https://xcorr.net/2011/05/27/whiten-a-matrix-matlab-code/
• http://eric-yuan.me/ufldl-exercise-pca-image/

Hoping it will be helpful

Dear Ali Mir,

Look the link, may be useful.

http://www.ijcim.th.org/past_editions/2008V16N2/P1-IJCIM16n2-May-Aug-2008-Improved%20Neural%20Network%20Performance-IJCIM-p1-8.pdf

Regards, Shafagat

Dear my friends,

I really appreciate your help.

Hi,

Firstly, apply PCA on your dataset X :: [COEFF,SCORE,latent] = princomp(X);

Then, pick the first k out of p columns of the SCORE matrix as your new reduced data. To choose k, execute this command: cumsum(latent)./sum(latent); It returns a vector by length of the p like as: 0.85, 0.92, 0.95, 0.99, ..., 1

Then if you want the reduced data covers at least 95% of the variance of the original data, you should set k = 3. (https://www.mathworks.com/help/stats/princomp.html)

Afterwards, you can fit an ANN algorithm on the reduced data:

https://www.mathworks.com/help/nnet/gs/fit-data-with-a-neural-network.html

Though your intention is fully applicable, I would like to provide another view to improve your idea. 

PCA is a method to reduce the dimension of input data itself, but it could also be treated as a flow of information given limited sensor/capacity of channels.

It is straightforward to replace the traditonary PCA process(singular value decomposition of the covariance matrix, etc...) with a naive linear neural network. Say we have N input units and N output units, we set only one layer of linear latent units with M units with M < N. And we train a function of 'equal' on this simple neural network. The error function is given by the least square. 

It is easy to prove that the function given by this ANN equals a linear transformation given by AB, where A is a N*M matrix and B a M*N one, it is easy to know AB has rank M, and therefore it provides the effect of PCA.

This approach can be applied to the lower layers of an ANN with appropriate pre-training. After pre-training, the outputlayer of this simple can be removed without loss of information. 

The idea that delicate and well-defined PCA is actually a special case of ANN may be a little intriguing and annoying, but I hope it would help your project in one way or another.

Yours Sincere.