improve pattern recognition by neural netowrk

hello i have a feature vector localbinary pattern histogram as Nx400 or 400 samples .I have normalised each of the columns of this matrix using the formuala=(value=value-mean/std deviation). The matrix has values like this:

    92    0.2363   -0.8730   -0.2689    0.4363   -0.8475    0.4239   -0.1029   -0.4275    1.4160    1.6939    0.4675
   -0.3125    1.7340   -0.9160   -0.2689    0.3944   -0.7625    0.4239    0.0572   -0.5116    1.6515   -0.4644    0.4675
    0.3799    0.7462   -0.1863   -0.2689    0.3944   -0.9750   -0.9408    1.4239   -0.9741    0.8166    0.0434    0.4254
    1.8187    0.0876   -0.8730   -0.2689    1.0639   -0.2950    1.0856    1.7442    0.0350    1.7157    1.6833    1.0998
    1.0722    0.1407   -0.9160   -0.3550    1.0639    1.5857    0.3825    0.6551   -0.3013   -0.5212    1.7257    1.1420
    1.7646    0.4806   -0.2292   -0.5705    1.3986    1.5007    1.0856    1.6694    1.3910   -0.8423   -0.3903    0.4254
    1.7213    0.4806   -0.2292    0.0650    1.6497    1.2351    0.4239   -0.3165    1.4961   -0.6925   -0.6019    1.0998
   -0.5721    0.4169   -0.2292    1.7022    1.6497    1.5539    0.5066   -0.6795    1.6853   -0.8637    1.4294    1.1420
    1.5374    1.0967    1.1444    1.8314    1.6497   -0.6562    0.4239   -0.6795    1.0021   -0.8423    1.0909    0.4254
    0.4989    1.0861    1.1766    0.6466    1.6915   -0.8262   -0.9408   -0.6582   -0.3434   -0.6604    1.4294    1.4371
    1.1912   -0.2417    1.1766    1.7991    1.0011   -0.8262    1.0856   -0.7009   -0.4695   -0.8423    1.3448    1.7321
   -0.9183   -0.9534   -0.9160   -0.9151   -0.9758   -0.9750   -0.9718   -0.9785   -0.9741   -0.9922   -0.9722   -0.9552
   -0.9183   -0.9534   -0.91

I want to train the neural netowrk to do multi class classification to perform face recognition u the eror is very high on the test set and also it is doing lot of mis classifications. How to improve the performance and hence generalization.I am attaching my code below.Please help!.If I am doing any fundamental mistake let me know!!

P.S The way I am doing my indices matrix(ind) in my code below is based on the fact that the feature vector is extracted from a imageset which which has 40 classes s1 s10 s11...s40. Hence i am repeating each indices 10 times to display 10 images per class.Kindly see the code in the file attached below.

[Newman]

this is my matrix it has only 200 columns because it was exceeding 5 mb so...