neural network nprtool tansig vs logsig

Hello,

I am a little confused about the nprtool in the neural network toolbox. It generates a two layer feedforward network with a tansig activation on the output layer. However it expects a binary output with {0,1} and it seems to work right.

I wonder why doesn't it use logsig activation if the output will be {0,1}. When I manually modify the output activation to logsig the generated output gets compressed to [0.5,1] range, which is wrong.

I can't explain what seems to be the problem.

Thanks

Answer 1

[Greg Heath]

Most of the nets have a default mapminmax processing default that transforms outputs to the range (-1,1). For that range tansig is appropriate. There also is a default reverse transformation that yields the original target range.

My guess is that when you imposed 'logsig' the processing got screwed up.

Exactly what commands did you use to switch to 'logsig'?

Hope this helps.

*Thank you for formally accepting my answer.

Greg

Answer 2

[Greg Heath]

You can not manually add logsig after being trained with purelin

 logsig(0:1) = 0.5 0.7311

The net has to be trained with logsig.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Cagdas Ozgenc]

No I train with logsig by generating a script first from nprtool and modifying the activation function before running the script.

[Greg Heath]

Please post the script , initial RNG state, and the results from using one of the MATLAB classification nndatasets.

Answer 3

[Vito]

This Fuzzy logic. AND = min(a,b), OR=max(a,b). The binary operator S can represent the addition(OR) boundary: S(1, 1) = 1, S(a, 0) = S(0, a) = a (logsig)

[Greg Heath]

Wha??

[Vito]

Look at the theory, since classical logic, three-value and more fuzzy logic. The base of fuzzy logic is the algebra of Minima and Maxima which has the same properties, as Boolean algebra. In the help of ML about it it is told in general.