improve accuracy in pattern recognition using neural network

Hello I am trying to build a face recognition system using neural network. For that I am using SURF features as a feature vector for the input. Here is my code:

%%extract the face databse from the folders

                imdataloc='D:\xdf\orl_faces';
           facedatabase=imageSet(imdataloc,'recursive');

          fprintf('The number of unique individuals are %d\n',length(facedatabase));

              sd=input('enter the individual under consideration');

%display montage for the individual under choice

                 montage(facedatabase(sd).ImageLocation);

%%create a gallery of all individuals

         for i=1:length(facedatabase)

           image(i)=facedatabase(i).ImageLocation(1);

         subplot(1,2,1)
         imshow(read(facedatabase(sd),1));
           subplot(1,2,2);
          montage(image);

            end

% split the database in to 70 20 split

       [training,testing]=partition(facedatabase,[0.8 0.2]);

%extracting SIFT features fromeach image of the training data

      extractorFcn = @exampleBag;% where example bag is SURF
      bag = 

       bagOfFeatures(training,'CustomExtractor',extractorFcn,'Vocabulary',480);
      bagtest=bagOfFeatures(testing,'CustomExtractor',extractorFcn );

    featureVectortest =(encode(bag,training))';% the input feature vector

    ind=[1 10 11 12 13 14 15 16 17 18 19 2 20 21 22 23 24 25 26 27 ...
    28 29 3 30 31 32 33 34 35 36 37 38 ...
    39 4 40 5 6 7 8 9];
     ind = repelem(ind,8);% repeat the individual columns 8 times to represnt 80 percent of the dataset images in the training set

         target = full(ind2vec(ind));% the target matrix

%save('SIFTneuraldata.mat','featureVectortest','target','training','testing','bagtest');%ssave the input and

       for i=1:5
     imagetest{i}=read(training(i), 1);
  end
    for j=1:5
     featurevectorpar{j}=(encode(bagtest,imagetest{j}))'; 
    end

%this is the query feature for 5 images from the test set

     featurevectorpar=cell2mat(featurevectorpar);

training is done using patternet where net.divideParam.trainRatio = 100/100; as I am splitting the database already.The databse I am using is AT&T cam bride face dataset 40 individuals and 10 images per individual

    The image set has been formed in this manner :
    Image set 1: s1.
   Image set 2: s10.
  Image set 3: s11.
    Image set 4: s12.
   Image set 5: s13.
   Image set 6: s14.
   Image set 7: s15.
   Image set 8: s16.
    Image set 9: s17.
    Image set 10: s18.
    Image set 11: s19.
   Image set 12: s2.
   Image set 13: s20.
   Image set 14: s21.
   Image set 15: s22.
   Image set 16: s23.
   Image set 17: s24.
   Image set 18: s25.
   Image set 19: s26.
   Image set 20: s27.
   Image set 21: s28.
  Image set 22: s29.
  Image set 23: s3.
  Image set 24: s30.
   Image set 25: s31.
  Image set 26: s32.
  Image set 27: s33.
  Image set 28: s34.
  Image set 29: s35.
  Image set 30: s36.
  Image set 31: s37.
  Image set 32: s38.
  Image set 33: s39.
  Im *age set 34: s4.
  Image set 35: s40.
  Image set* 36: s5.
  Image set 37: s6.
  Image set 38: s7.
  Image set 39: s8.
    * Image set 40: s9.

So in that manner I have made the class indices above. I am also restricting vocabulary of training set to be 480 and not default 500 because the test set is giving me 480 so feature vector has to be of consistent size .

*but the problem is tht the performance is very poor I dont know why perhaps because of less images or because I am labeling indices wrong in the target matrix.Kindly help me on how to improve this recognition *