MATLAB: Bad results of the neural network _ newsgroup

Deep Learning Toolboxneural network

Hi Dr Greg i tried to reply to your message in the NEWSGROUP in this link several times https://www.mathworks.com/matlabcentral/newsreader/view_thread/348826 but but it dosen't post , i don't know why so i decide to post it here as a question because i couldn't have any solution .
This is the message that i wanted to post :
I appology for that DR Greg this will not be repeated
> x = patientInputs;
> t = patientTargets;
> inputs = mapstd(x); 
> targets = mapstd(t); 
>> I don't remember suggesting mapstd. Data is automatically normalized 
and unnormalized by the training function. The only reason to use mapstd 
is to detect and delete or modify outliers.
I asked you because of the format of my inputs as you can see in this link https://www.mathworks.com/matlabcentral/answers/344813-how-to-do-when-the-inputs-ranges-for-neural-network-are-not-so-uniform-in-magnitude Actually I tried to use MAPSTD but it dosen't improve so i applied the logarithme to the inputs that have a very high value numerically compared to other inputs but also it doesn't work .
> N=981
>>Why does this differ from the 1012 of previous posts?
It was at first 981 but i wanted to increase the number of N to see if this will improve the peformance and as i see that there is no big progress i returned to 981
> Ntrn = N-2*round(0.15*N) % 

> Ntrneq = Ntrn*O % 
>>If you are not going to use an ending semicolon, then print the answer
after the percent sign. For example
[ I N ] = size(x) % [ 9 981]
[ O N ] = size(t) % [ 2 981 ]
Ntrn = N-2*round(0.15*N) % 687
Ntrneq = Ntrn*O % 1374
> %For a robust design desire Ntrneq >> Nw or
> H=10
or H = 10 ???
>>That doesn't make any sense
This was just a typing error
 > Hub = -1+ceil( (Ntrneq-O) / (I+O+1)) % Hub =117
>>117?? Your arithmetic sucks.
> Ntrials = 10
> rng(0)
> j=0
> for h =round([Hub])
>>Where in the world did you get that from? Hub = 114 !
> j = j+1
> h = h %12
>> h = 114! Where did you get 12 from??
Actually i got this value after running the code and it was from the given results
> trueclasses = vec2ind(t);
> for i = 1:Ntrials
> net = configure(net,x,t);
> [ net tr outputs errors ] = train(net,x,t);
> assignedclasses = vec2ind(outputs);
> classerr = assignedclasses~=trueclasses;
> Nerr(i,j) = sum(classerr);
> % FrErr = Fraction of Errors (Nerr/N) 
> [FrErr(i,j),CM,IND,ROC] = confusion(t,outputs);
> FN(i,j) = mean(ROC(:,1)); % Fraction of False Negatives
> TN(i,j) = mean(ROC(:,2)) ; % Fraction of True Negatives
> TP(i,j) = mean(ROC(:,3)); % Fraction of True Positives
> NMSE = mse(errors)/mean(var(t',1)) 
> end
> end
> PctErr=100*Nerr/N
> NMSE = mse(errors)/mean(var(t',1)) 
>>This makes no sense: You have Ntrials designs but only take 
the last one instead of the best one
the original code that i got is from you answer in this link https://www.mathworks.com/matlabcentral/answers/130271-how-to-improve-accuracy-for-unseen-data
% NEURAL NETWORK MODEL
but i tried to add NMSE = mse(errors)/mean(var(t',1)) and it was a fault from me to put it after the loop.
My real problem is how can i use this result to improve the performance of my neural network So as i understand that for Not overfitting Ntrneq >= Nw so i can have idea about the limit of H because Nw = 12*H+2 and i should decrease H to have NMSE <= 0.01 * MSE00 . Ami right ?
Again i apology for this bad Etiquette but i hope that you understand me because i have no idea about neural network and Matlab and i have no time for my project this is why .
Thank you Dr greg for you patience .

Best Answer

1. You did not define net = patternnet
2. Even though it is a classifier, just try to minimize H subject to the constraint
    mse(e) <= 0.01*mean(var(t',1))  % e=t-y
i.e.
     NMSE <= 0.01
or
     Rsquare >= 0.99
Hope this helps
Thank you for formally accepting my answer
Greg
3. If you overfit, validation stopping will kick in
Good Luck
Thank you for formally accepting my answer
Greg
Related Question