How to use positive part: subplus(x) for neuron model (instead of tansig(x) or logsig(x))?

Hello,

I want to use nonlinearity of subplus(x) = [x]+ and learn a neural network with back propagation. I understand that i must use a differentiable function. i can use instead the function y= log(1+exp(x)). I have 2 questions:

1. How can i create a transfer function such as y = log(1+exp(x)) instead of logsig or tansig?

2. Is there another way to do that? thanks a lot, Elia

Answer 1

[Greg Heath]

In the command line:

Use the commands

 lookfor, help, doc and type

on the functions

 hardlim
 perceptron
 adapt

The percepton uses the step transfer function hardlim. Hardlim is not differentiable. Therefore backpropagation via train cannot be used. Instead, use the learning function adapt.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Greg Heath]

I don't understand why anyone would want to use subplus in a NNET.

Will you please explain your reason?

By the way subplus is differentiable. It's derivative is hardlim.

[Rena Berman]

(Answers dev) Unaccepted answer due to answerer's flag.