which is the best neural network for classification problem?

for 2-class classification problem, which is the best neural network, what transfer function to use and the number of desired number of neurons?

[Image Analyst]

I imagine that Greg would say "It depends, and you have not supplied enough information to even give an answer", so I would recommend you read this and fix your post.

Answer 1

[Greg Heath]

It depends.

However, in general, the default should be the standard classification version of the universal approximator. The MATLAB version is PATTERNNET.

 help patternnet
 doc patternnet

For practice data

 help nndatasets
 doc nndatasets

For nontrivial examples search the NEWSGROUP and ANSWERS using

 greg patternnet

Hope this helps.

Thank you for formally accepting my answer

Greg

[Dhandapani.S]

i am working on handwritten signature verification modelling. i extract features of the signature andtryinh to use neural network for clasification. wht should be the target value. i have ketp 1 for genuine and 0 for forged signatures. what will be the best training trsting ratio?

[Greg Heath]

 Total      = design + test
design      = training + validation 
nontraining = validation + test

The ratio is not important. The most important thing is to have enough training examples to design an accurate net that is robust with respect to noise, interference and transcription errors.

For an I-H-O node topology

 Ntrneq = Ntrn*O            % No. of training equations
 Nw     = (I+1)*H+(H+1)*O   % No. of unknown weights

For robustness desire

 Ntrneq >> Nw    <==>  H << -1+ceil( (Ntrneq-O) / (I+O+1) )

Otherwise consider validation stopping and/or regularization (msereg and/or trainbr).

Hopefully, there is enough left to have reasonably precise estimates on

 a. nontraining test data
 b. nondesign validation data

If not, you can resort to regularization instead of validation and/or multiple crossvalidation for precise test estimates.

Typically, my training goal is

 Minimize H subject to mse(target-output) < 0.01*mean(var(target',1))

Then the net successfully models 99% of the target variance.

Hope this helps.

Greg