Manual Calculation of Neural Networks

Is there a way to manually follow the equations on a neural network that has been trained and tested using the ANN toolbox?

I want to see if I can follow the equations and see where my app that I've created that replicates the values and calculations of my trained network is differing.

Thanks,

C

Answer 1

[Greg Heath]

If the classes are not mutually exclusive (e.g., tall, dark, handsome) logsig is an appropriate output transfer function for approximating input-conditional posterior probabilities.

For mutually exclusive classes, the probabilities sum to one. The unit-sum softmax is appropriate.

Answer 2

[Greg Heath]

Follow? Not exactly. However you can reproduce the results. For FITNET,

 yn = b2+ LW * tanh( b1 + IW * xn );

where

 xn and yn are the normalized input and output.

What net and options are you using?

[Conor O' Sullivan]

I was originally using the Pattern Recognition NN Wizard from the toolbox. It worked really well but I need to replicate the "math" in an app I'm developing. However, the built in softmax transfer function is beyond me, so I want to try and replicate this network but with tanh functions and with hardlim functions. I'm still struggling to do that though, so I was hoping someone more knowledgable might have an answer...