Prev: waitbar for parfor loops
Next: pitch detection
From: Ali on 3 Nov 2009 15:22
hi guys,
Is it better to use 'logsig' or 'tansig' for transfer function of
output layer? is it important?


if we use tansig, can we use output data in [0 1] range (normalize)?
From: Greg Heath on 4 Nov 2009 22:22
On Nov 3, 3:22 pm, Ali <vahed.k...(a)gmail.com> wrote:
> hi guys,
> Is it better to use 'logsig' or 'tansig' for transfer function of
> output layer? is it important?
>
> if we use tansig, can we use output data in [0 1] range (normalize)?

It depends on the output scaling.

For regression with no natural finite bounds (e.g., [0,2*pi) for
angles)
I standardize the outputs (zero mean and unit variance) and use
PURELIN.

For regression with natural finite bounds (e.g., [0,2*pi) for angles)
I normalize the outputs to [-1,1] and use TANSIG. Also used for
bipolar binary representations.

For classification the class targets are coded using columns of the
unit matrix so that outputs represent posterior probabilities
estimates.
LOGSIG is the proper choice for this scenario and , im general,
for unipolar binary inputs.

See my post on pretraining advice:

http://groups.google.com/group/comp.soft-sys.matlab/msg/0d24fcb92959575a?hl=en


Hope this helps.

Greg
 | 
Pages: 1
Prev: waitbar for parfor loops
Next: pitch detection