I don't Know why my neural network doesn't give good results

I tried to build NN for spoken word classification I followed many approaches discussed here in th group, yet the results terrible I got only ~65% correct classification, I don't what wrong I'm so desperate.

I appreciate any help or notice on my code or the approach I followed

load yesClass2;
load noClass2;
yes2=1*ones(1,199);
no2=zeros(1,208);
InData=[yesClass2 ;noClass2];
InData=InData';
TarData=[yes2 no2];
xtrn=InData;
ttrn=TarData;

[ I N ] = size( xtrn ) % [ 5 407 ]
[ O N ] = size( ttrn ) % [ 1 407 ]

MSEtrn00 = mean(var(ttrn',1))
MSEgoal  = MSEtrn00/100
MinGrad  = MSEtrn00/300

%rng(0)
net4 = patternnet(58,'trainlm');
net4.divideFcn = 'dividetrain';
net4.trainParam.goal = MSEgoal;
net4.trainParam.min_grad = MinGrad;
[net4 tr ] = train(net4,xtrn,ttrn);
bestepoch = tr.best_epoch;
R2 = 1 - tr.perf(bestepoch)/MSEtrn00;
save net4 net4

1. I chose No. of hidden nodes 58 based on max. R2 achieved
2. max. R2 = 0.9
3. I attached confusion matrix for complete data set used for training and for divided data set into 70% training, 30% validation and 30% testing

[Sarah Mahmood]

I forgot to mention

MSEtrn00 =
    0.2499
MSEgoal =
    0.0025
MinGrad =
   8.3293e-04

This is the first confusion matrix

and this is the second confusion matrix.

[Greg Heath]

Totally confusing post

=====================

Came back later and struggled through. I understand now.

Next time you post code please explain exactly what you are doing and why.

In particular 58 hidden nodes is overfitting when the data is divided.

Unfortunately, the MATLAB confusion matrices are not very easy to understand.

I have posted a better code for understanding (might be in the NEWSGROUP).

Greg

Answer 1

[Greg Heath]

This is a classic case of overtraining an overfit net:

DIVIDETRAIN:

 Nw = (5+1)*58+(58+1)*1 = 348 + 59 = 407 % Unknown weights

 Ntrneq = 407*1 = 407 % Equations

The no overfitting condition Ntrneq >> Nw is readily violated

DIVIDERAND:

 Ntrneq = 407 -2*round(0.15*407) =  285 < 407

[KAE]

Here are Greg's calculations with extra comments to help other learners,

% Number of rows: number of features. Number of columns: number of samples.
[ I N ] = size( xtrn ) % Size of network inputs  
[ O N ] = size( ttrn ) % Size of network targets

% H is number of neurons in hidden layer, here 58

fractionTrain = 0.15; % Fraction of data used as training examples. 
% 0.15 is the default, which is assumed here
fractionValid = 0.15; % Fraction of data used as validation examples
% 0.15 is the default
Ntrn = N - round(fractionTrain*N + fractionValid*N); % Number of training examples
Ntrneq = Ntrn*O; % Number of training equations
Nw = (I+1)*H + (H+1)*O; % Number of unknown weights
% Since we want Ntrnreq>>Nw, we could require that Nw<Ntrneq/10
% But it's not, so we are overfitting
% Must either get more data (training examples), or 
% simplify our model (smaller H)

The plot here is helpful for understanding overfitting.