I DO NEED HELP, problem with number of neurons in hidden layer of feedforward neural network

I do need some one to follow my code, i have attached my m-file here! I don’t know what the hell is wrong with my code whatever I do it works too good to be correct My neural network, no matter how many neurons I put for its hidden layer it works too good , I know there must be something wrong , if u can figure out where I am wrong , please let me know ,I got so frustrated by this program , what I am doing is a simple classification of the glass data which I got from here: http://archive.ics.uci.edu/ml/machine-learning-databases/glass/ I try to classify this data into 7 classes, no matter how many neurons I have in hidden layer the training and learning errors are about 0. How is it possible? I know I must be wrong somewhere, but I cannot figure it out:/

please help me with this!

if true
  %NET STRUCTURE

net=feedforwardnet(1);% i have changed number of neurons here

% but no matter what it is , it works %,i mean the learning error and

%test of error are close to zero:/

net.divideParam.trainRatio=0.9 net.divideParam.valRatio=0.1 net.divideParam.testRatio=0

%LEARNING PHASE

[net,tr]=train(net,OL,traininglabel); net.trainparam.epochs=1000; yt=net(OL); E_oL=mse(yt-traininglabel) learnigErrorofactualdata(ii,1)=E_oL;

%testing PHASE BY USING 20% OF ACTUAL DATA

kv=testdata'; Dv=testlable';

Ytest=net(kv); E_ov=mse(Ytest-Dv);

ValErofACtudata(ii,1)=E_ov end

i attached my m-file , please have a look on it ,thanks in advance , i do need help

[Image Analyst]

Read this or just attach your m-file with the paper clip icon.

[Fereshteh....]

i just attached my m-file

[Greg Heath]

It would be preferable for you to

 1. Post your code and initial RNG state
 2. Post the results of running the code on one of the MATLAB classification examples obtained from

     help nndatasets
     doc nndatasets

Hope this helps.

Greg

[Greg Heath]

The glass_dataset in the MATLAB nndataset has the same input matrix, as the one in http://archive.ics.uci.edu/ml/machine-learning-databases/glass

However, the target matrices are different:

 MATLAB:  Two target classes: windowpane and nonwindow 
 UCI.edu: Seven target classes: Seven window subclasses

Using the MATLAB data,

 I determined that the 2 categories provided more equations than unknown weights when the umber of hidden nodes, H, was less than the upper bound Hub = 35 . So I first did a broad search with h = 0:2:18 = ceil(Hub/2). Then, seeing that H = 4 was sufficient, I fine-tuned the search to h = 1:4 and found that 3 hidden nodes is sufficient.

If I assume that you used the default value of H = 10 but had 0.8*7/2 = 2.8 times as many training equations, I'm not surprised that performance was good.

I cannot follow your code. However, I suggest that you try to minimize the number of hidden nodes. This will increase the robustness of the design wrt generalizing to unseen nontraining data.

Hope this helps.

Greg

[Fereshteh....]

thanks Greg , the thing is no matter how many neurons i put in hidden layer, even one neuron :O it works too good , i must be wrong somewhere i don't get where , i need someone to have a look on my net structure to see if i wrote it properly

[Greg Heath]

>> a=importdata('E:\all uni stuff\thesis\data-wine-breast-\glasses.txt') Error using importdata (line 137) Unable to open file.

>> "http://archive.ics.uci.edu/ml/machine-learning-databases/glass" "http://archive.ics.uci.edu/ml/machine-learning-databases/glass" | Error: The input character is not valid in MATLAB statements or expressions.

The code you posted DOES NOT RUN when cut and pasted into the command line. It would also be helpful to remove the ending semicolons on SELECTED lines so that nonvoluminous intermediate results are printed to the screen

Greg

[Fereshteh....]

Greg i just fixed my problem i guess, thank you very much , you did help me , your first comment was very helpful ,thanks in count of stars in sky

[Greg Heath]

Please explain and post your corrected code.

Did you do anything special to deal with the empty class 4?