Neural Network program problem in classification

Hi All

I am using this code to train my network, the problem is , if I give an input that is somehow among the value of the inputs I have chosen to train , it gives the right output , but if I give something out of this range , still the output is in the same range of the targets I have given to the code :

close all, clear all, clc, plt = 0
load('input.txt')
%load input 
load ('target.txt')
%normalizing data 
input=input';
target=target';

% input =  mapstd(input);
% target =  mapstd(target);

x=input;
t=target;

% [ x, t ]  = simpleclass_dataset;
[ I   N ] = size(x)   % [ 2 1000 ]
[ O N ] = size(t)    % [ 4 1000 ]

%vec2ind Transform vectors to indices. takes an NxM matrix V and returns a 1xM vector of indices
% indicating the position of the largest element in each column of V.
trueclass = vec2ind(t);  

class1 = find(trueclass==1);
class2 = find(trueclass==2); %in my example all the largest elements are on the 2nd column
class3 = find(trueclass==3);
class4 = find(trueclass==4);

N1 = length(class1)    
N2 = length(class2)   
N3 = length(class3)    
N4 = length(class4)    
x1 =  x(:,class1);
x2 =  x(:,class2);
x3 =  x(:,class3);
x4 =  x(:,class4);
plt = plt + 1
hold on
plot(x1(1,:),x1(2,:),'ko')
plot(x2(1,:),x2(2,:),'bo')
plot(x3(1,:),x3(2,:),'ro')
plot(x4(1,:),x4(2,:),'go')
% 
% Nw = (I+1)*H+(H+1)*O;

Hub = -1+ceil( (0.7*N*O-O)/(I+O+1))  % 399
Hmax = 40   % Hmax << Hub 
dH    = 4   % Design ~10 candidate nets
Hmin  = 2   % I know 0 and 1 are too small
rng(0)   % Allows duplicating the rsults
j=0
for h=Hmin:dH:Hmax
    j = j+1

      net            = patternnet(10);
      net = init(net);  % Improving Results since we use patternet we should use init
      [ net tr y  ]  = train( net, x, t );
      assignedclass  =  vec2ind(y);
      err = assignedclass~=trueclass;
      Nerr           = sum(err);
      PctErr(j,1)    = 100*Nerr/N;
  end
  h = (Hmin:dH:Hmax)'    
  PctErr = PctErr

I just want to know , according to the graphs of confusion , performance ,and the classes drawn , is the training enough or too much or little ?

Answer 1

[Greg Heath]

You are confused.

What is the physical problem you are trying to solve?

1. What are the inputs?

2. What are the corresponding outputs?

3. The targets are not unit vectors. Therefore they are not appropriate for use in a standard NN classifier.

[farzad]

Dear Professor

My problem is something related to a structure, and If I understood you correctly ,it is not an image processing , so the order of the numbers are as in the files , also , each sample data are written in each row, so to use it in Matlab ,I have to change the rows with columns

the Targets can not be unit vectors in their actual value , and I don't know if one can change it to unit vectors

so , my question is : Can we use Neural Networks , classification , for a Problem with the outputs that are not only 0 and 1 ??

I add that I did not Accept this answer and I do not know how your answer was accepted , but I am not yet convinced and yet not have got my answer , in your previous answer , I put the code according to the MATLAB HELP EXAMPLE , but , my data all go to the same class !!

[farzad]

for the questions :

1- the inputs are 789 Samples of 8 elements

2- the outputs are 789 Samples of 3 Elements

3- in this case , my question is : Can my problem be solved by NN at all after being transfered to a unit vector, or there is no such possibility and solution to my problem ?

[Greg Heath]

% My problem is something related to a structure,

 Uninformative

%and If I understood you correctly ,it is not an image processing ,

 When did I imply that?

% so the order of the numbers are as in the files ,

 I don't know what you mean.

% also , each sample data are written in each row, so to use it in Matlab ,

% I have to change the rows with columns

 OK

% the Targets can not be unit vectors in their actual value , and I don't

% know if one can change it to unit vectors

 Then how are you going to design a useful classifier?

% so , my question is : Can we use Neural Networks , classification , for a

% Problem with the outputs that are not only 0 and 1 ??

 Yes. Provided you know what you are doing. However, you do not. So, my 
advice is 

 If you have c classes , convert the columns of the target to columns 
of eye(c) with the row index of the "1" indicating the class index in {1:c}

% I add that I did not Accept this answer and I do not know how your answer

% was accepted , but I am not yet convinced and yet not have got my answer

 OK. I flagged the acceptance (don't know how to remove it).

% , in your previous answer , I put the code according to the MATLAB HELP

% EXAMPLE , but , my data all go to the same class !!

 If your targets are not unit vector columns then what you did makes no sense.

%3- in this case , my question is : Can my problem be solved by NN at all

%after being transfered to a unit vector, or there is no such possibility

%and solution to my problem ?

It depends on

(1) the data (2) what you mean by solved ... minimize error rates on nontraining data?

[farzad]

Dear Professor Heath

the input is some material properties, the output are some contstants and dimensions , so they are not dimensionally related to each other , I really don't know how should I define my problem to help more in understanding the problem.

Still , I have tried to convert the target matrix to the way you have mentioned , if I have understood correctly , and I attached the new files , but no iterations goes on , and this time nothing happens !!!

[farzad]

by being solved I mean , for example if I give the inputs :

input = [20793635.0022993088 3308122.1690468886 484.8760825559 1622.0869505062 1.9257234725 0.8661640625]

I should get the output (as a test) :

output= [0.0010000000 35.0000000000 0.0600000000]

as you see , the second element is always the largest .

[farzad]

I have searched all the web , never found a similar example in classification problem , that the indice of the vectors is always one number

[farzad]

Dear Professor I think I have managed to create to appropriate file , If I am not wrong , but now I have a problem with what that is happening with matlab when starts to train the network , the GRADIENT is NaN , and the performance is to infinity , and not accessible , could you please help me with this problem

Answer 2

[Brendan Hamm]

Just looking at this briefly, you have multiple output classes but only 1 class that is being used for training data. Therefore there is no distinction for the classification to make ... everything is of the same class. If you train me to classify everything I see as a circle, and then give me a square, I classify it as a circle still. You need a better training set or your neural net is pointless.

[farzad]

Thank you dear Brendan , This was Exactly my question , also in my previous questions , thank you , ok , I am searching in the posts , to understand the classification , or find a method , but first I don't really find a good example , for my case , and I don't know the strategy , like the first photo I have inserted , how could it help me ? and if there is any suggestion for distinguishing , please help me , thanks a lot

Answer 3

[Greg Heath]

The code you are using is for 4 classes.

Revise the code for the number of classes that you have,

For c classes the columns of target are {0,1} c-dimensional unit vectors.

[farzad]

Thank you dear professor heath

Yes I have to change the classes to 8, but my question is how to write that part to not to put all the input in one class, because in the code using the vec2ind leads to assignment of all the inputs to one class

[Greg Heath]

For classification of c classes the columns of the target are columns of eye(c) with the 1 in the row corresponding to the corresponding class index.

If you do not understand that, look at the target matrices for the classification examples in the documentation

 help nndatasets

[farzad]

Dear Professor

the examples only show how the matrices are , but the Target data are already prepared in the simpleclass -dataset

but this does not Answer my question , because I have a target matrix , that I need to know , how to convert it to the {0,1} c-dimensional unit vectors , because the vec2ind(target) ONLY finds the Largest element in each row of the target , that in my case it's the second column in all of the rows , so the trueclass matrix in my case is a row matrix that all the elements are 2 , the code doesn't give me such matrix as you have mentioned

[farzad]

here is the example of classification in nndatasets :

[x,t] = simpleclass_dataset;
    plot(x(1,:),x(2,:),'+')
    net = patternnet(10);
    net = train(net,x,t);
    view(net)
    y = net(x)
    plotconfusion(t,y)

it just loads the simpleclass_dataset , that the {0,1} c-dimensional is already made , I don't know how !!!

Also I don't know how the second line of the above code help us ? what does that plot say ??!!

[farzad]

I have tried the example of nndataset for my case , but yet : it puts all the targets at the same class so they are accumulated in class 2 , according to the confusion matrix , so I think , for my type of data , I need some other comment , or there is a missing comment

close all, clear all, clc, plt = 0
load('input.txt')
%load input 
load ('target.txt')
%normalizing data 
input=input';
target=target';

% input =  mapstd(input);
% target =  mapstd(target);

x=input;
t=target;

setdemorandstream(491218382)

net = patternnet(10);
view(net)

[net,tr] = train(net,x,t);
nntraintool

plotperform(tr)

testX = x(:,tr.testInd);
testT = t(:,tr.testInd);

testY = net(testX);
testIndices = vec2ind(testY)

plotconfusion(testT,testY)

[c,cm] = confusion(testT,testY)

fprintf('Percentage Correct Classification   : %f%%\n', 100*(1-c));
fprintf('Percentage Incorrect Classification : %f%%\n', 100*c);

plotroc(testT,testY)

[farzad]

 my target vector should look like a [0 1 0;0 1 0;0 1 0 ;........;0 1 0] but I see that the lines of the example code , do not do this for my case

[farzad]

Also , Talking about the first code , there is no difference between the input vector x and x2 at all !!! my question is , if all the data go to the same class , is this correct ?