Neural Netwok testing issues

I am a beginner on neural networks and working on a project in which I am supposed to recognize a face of a person. I have developed a code and train the network on NN toolbox with the specific data base(5 people each having 10 samples).

During test when I provide one of the picture from data base(through which I have trained the network) it works correctly( gives me 1 or 0.99 value in output on the specific person)

But my problem is this when I provide it a random pic whose dimensions are same it never gives zero output for all the faces(those 5).

Answer 1

[Greg Heath]

Bottom lime:

You don't have enough data to adequately characterize a space of dimension 10,304.

50 data points can span, AT MOST, 49 dimensions.

Use feature extraction to drastically reduce input dimensions.

Try searching the NEWSGROUP and ANSWERS for ways to extract features from face images

Hope this helps.

Greg

Answer 2

[Greg Heath]

Are you using patternnet? If not, what?

What is the size of your input matrix, target matrix and number of hidden nodes?

What defaults have you overridden?

Your target columns should be 5-dimensional unit vectors with 1 unity component and the rest zeros.

trueclass = vec2ind(target) % target = ind2vec(trueclass)

Your output columns should be 5-dimensional consistent estimates of class posterior probability unit vectors with the row index of the largest component indicating the assigned class. It is not necessary that the rows have unity sums. However, if you wish to quote posterior probabilities, just divide each output column by it's sum.

assignedclass = vec2ind(output)

The most common reasons for poor performance on nontraining data are

 1. The training samples are nonuniformly distributed among the classes
 2. Nontraining samples from one or more classes are not sufficiently 
characterized by the training samples
 3. The net is overfit with too many weights that are not sufficiently constrained by the training samples.

Hard to say more without more info and/or posted code.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Mehrukh Kamal]

http://mehrukhkamal27.tumblr.com/

variable I has a picture which is not present in the data base but output 5 is 0.9551. Why isnt it near zero or still a less value?

Answer 3

[Mehrukh Kamal]

Thank you so much for your response.

Yes i'm working on pattern recognition.

input_matrix_size=10304x45

target_size=5x45

we have created the matrix in the same manner like you have mentioned. the samples are uniformly distributed as I have mentioned before each face is having 10 samples.

kindly explain this point :" The net is overfit with too many weights that are not sufficiently constrained by the training samples."

one thing I have to mention that I am directly working on code and have trained through NN toolbox I haven't gone through all these weight bias etc

[Greg Heath]

You have neither answered all of my questions, nor posted relevant code.

How do you expect us to help???

[Mehrukh Kamal]

http://mehrukhkamal27.tumblr.com/

variable I has a picture which is not present in the data base but output 5 is 0.9551. Why isnt it near zero or still a less value?

[Greg Heath]

Because the net thinks it is closest to class 5.

How can I confidently answer your questions when you haven't answered all of mine nor posted any code???

How can you have 10 examples each of 5 classes when you only have 45 total??

[Mehrukh Kamal]

Here's the code!

clear all close all clc fname=['D:\work\s']; for i= 1:5 b=num2str(i); fname1=[fname b]; for j=1:9 char='\'; pgm='.pgm'; c=num2str(j); d=[fname1 char c pgm];

imread(d); g=ans(:); h=g'; if i==1 if j==1 input_data=h; else input_data=[input_data;h]; end end if i>1 input_data=[input_data;h]; end end end input_data=input_data'; b1=0; %inital matrix to add further d1=ones(1,9); e1=zeros(4,9); target=[d1;e1]; for i1=1:4 starting_zeros=zeros(i1,9); ones_row=ones(1,9); rest_zeros=zeros(4-i1,9); s1=[starting_zeros;ones_row; rest_zeros]; target=horzcat(target,s1) end

[Mehrukh Kamal]

%%code 
clear all 
close all
clc 
fname=['D:\work\s'];
for i= 1:5
b=num2str(i);
fname1=[fname b];
for j=1:9
char='\';
pgm='.pgm';
c=num2str(j);
d=[fname1 char c pgm];
imread(d);
g=ans(:);
h=g';
if i==1
if j==1
input_data=h;
else
input_data=[input_data;h];
end
end
if i>1
input_data=[input_data;h];
end
end
end
input_data=input_data';
b1=0; %inital matrix to add further
d1=ones(1,9);
e1=zeros(4,9);
target=[d1;e1];
for i1=1:4 
starting_zeros=zeros(i1,9);
ones_row=ones(1,9);
rest_zeros=zeros(4-i1,9);
s1=[starting_zeros;ones_row; rest_zeros];
target=horzcat(target,s1)
end

[Greg Heath]

This code doesn't help w.r.t. your problem. It just shows that you read in the input matrix and created the target matrix. One of many simpler approaches:

 close all,clear all, clc
input_data = [];
target = zeros(5,45);
d1 =ones(1,9);
for i = 1:5
    for j=1:9
        g=rand(35,1);
        input_data = [ A g ];
    end
    j = 1+9*(i-1);
    target( i, j : j+8) = d1;
end

How can I help with the neural net code if you don't show any??

Answer 4

[Mehrukh Kamal]

How can you have 10 examples each of 5 classes when you only have 45 total??

 Sorry.. i actually amended the code and took 9 pictures for each sample and left one in each group for testing.

Answer 5

[Greg Heath]

What...no validation set? How did you avoid memorization by overtraining an overfit net?

[I N ] = size(input_data) = [10304 50 ]

[O N ] = size(target) = [ 5 50 ]

Neq = prod(size(target)= N*O = 250

==> Even if you used all of your data for training, you would only have 250 training equations to estimate ( for H = 10 hidden nodes)

Nw = (I+1)*H+(H+1)*O = 103050 +55 = 103,105

unknown weights!

You have to do some serious input feature extraction to reduce your input dimension to something much more reasonable.

See the comp.ai.neural-nets FAQ and posts regarding overfitting.

Even if you drastically reduce the input dimension, you should probably consider f-fold stratified cross-validation. For example, for each fold use, for each class, 8 for training, 1 for validation and 1 for testing. To get precise estimates of error rate you will have to make many runs. Keep track of running estimates of means and standard deviations of the validation and test error estimates. Do not include runs that do not converge to a reasonable estimate. If you plot the results (e.g., mean+/-stdv), you should see what is a reasonable threshold for exclusion.

Hope this helps.

Thank you for formally accepting my answer

Greg

Answer 6

[Mehrukh Kamal]

Thats the code.. there is no neural network code.. I have directly used NN tool box and provide the input and target matrix and trained it!

Answer 7

[Mehrukh Kamal]

What if i reduce the size of image to 50x50 then then matrics will be of 50x2500. will it be okey then?