Weight decay parameter and Jacobian matrix of a neural network

I want calculate prediction intervals so I have 2 direct questions:

1. How can I get the weight decay parameter 'alpha' (mse+alpha*msw) used when using 'trainbr' as a training algorithm?
2. How can I get the neural network jacobian matrix (derivatives following weights) calculated during training?

Answer 1

[Greg Heath]

The documentation for trainbr is pretty bad.

 help trainbr
 doc trainbr

Look at the source code

 type trainbr

I am not familiar with it but will take a look when I get time.

Meanwhile, if you make a run, the training record tr, contains 2 parameters

             gamk: [1x31 double]
             ssX:  [1x31 double]

that are involved.

Hope this helps.

Thank you for formally accepting my answer

Greg

Answer 2

[Platon]

Thank you. It helps to determine alpha but not to calculate the neural network jacobian matrix. Hope that future Matlab NN tool Box versions include specific tools to make prediction interval study.

[Greg Heath]

When using the obsolete msereg and mse with the regularization option, the weight parameters are alpha (specified error weight) and (1-alpha).

However when using trainbr, the weight parameters alpha and beta are calculated each epoch. Haven't decifered the logic yet. Might be faster to search the web.