Neural Network Classification: cost function, regularization parameter, availability of the hidden layers...

How do I derive the cost function J from a nn classification model?

How can I set a regularization parameter? Is this approached using your weights capability?

Are the hidden layer values available after modeling?

Answer 1

[Greg Heath]

You cannot derive a cost function from a model.

You specify the cost function for a design.

If you are new at this, start out by accepting the NNToolbox defaults. First check out the relatively simple examples in the classification documentation

 help patternnet

 doc patternnet

Then search the NEWSGROUP and ANSWERS for posted examples. For example

 greg patternnet

The default transfer functions are tansig (hidden layer) and softmax (output layer)

The "cost" function is crossentropy

Many questions on terminology and details can be obtained using the commands help, doc and type, e.g.,

 help patternnet

 doc patternnet

 type patternnet

Hope this helps.

Thank you for formally accepting my answer

Greg

[Richard Palmer]

Thanks for your response. I still have some questions. I believe that I know the formulae for the cost functions for the models in the toolkit. For instance, the cost function for a regularized LR function is

   (Sum(h(xi)-yi)**2 + lambda sum(theta**2))/(2 m)

for a regularized linear regression. I would like to retrieve the theta coefficients and want to know how to set the lambda (regularization) parameter.

The unregularized cost function will give different values for the training/evaluation/test sets. Can the crossentropy function be decomposed into it's constituent parts?

If this is in the documentation, I apologize. I haven't found the information at this point.

Again, thanks for replying.