Discussion:
size ANN training input
Andrea_Viano
13 years ago
Permalink
Hi,



Is there anyone that knows why training inputs of an ANN have to be less
than 1 to have a convergence? I tried with my regular input( 10000,or 8
etc), and I had a very high and constant error. If I divided training inputs
to reach numbers close to the unity I could find a solution of my ANN.



Can you help me to solve this problem?



Thanks



Andrea



Ing. Andrea Viano

Industry Division



_____

Logo_CFD_Engineering.jpg



CFD Engineering S.r.l.

Headquarters

Piazza della Vittoria, 7/2A - 16123 Genova (Italy)

Phone +39-010-8540050 - Fax +39-010-8631199
Cooper Stevenson
13 years ago
Permalink
Hi Andrea,

I have an algorithm I can provide you that will scale between 1 and 0. It's
written in shell but you're welcome to it if you like.


Very Best,


-Cooper
...
--
D. Cooper Stevenson
ph: 541.971.0366
em: cooper-/kSeq5SmJ+***@public.gmane.org
www: http://cooper.stevenson.name
Joan Maspons
13 years ago
Permalink
Hello,
Some activation functions outputs such as sigmoid ones give and output
between 0 and 1 or -1 and 1. If you try to get values around 10000 as a
pondered summation of the inputs and the outputs range is between 0 and 1
it's impossible to get such a big values. You can take a look to the
fann*scale* functions at
http://leenissen.dk/fann/html/files/fann_train-h.html.

J. Maspons
...
--
Joan Maspons
CREAF (Centre de Recerca Ecològica i Aplicacions Forestals)
Universitat Autònoma de Barcelona, 08193 Bellaterra (Barcelona), Catalonia
Tel +34 93 581 2915 j.maspons-KjRrbB+***@public.gmane.org
http://www.creaf.uab.cat
Fábio Blessa
13 years ago
Permalink
Normalization you mean?
Does it depends on the training algorithm? Having your data normalized is a
good practice.
...
Andrea_Viano
13 years ago
Permalink
Thank you Fabio,



I used a backpropagation method and I used activation sigmoid function.



Is it maybe for that I have to normalize?



Andrea













Da: Fábio Blessa [mailto:fabioblessa-***@public.gmane.org]
Inviato: mercoledì 2 novembre 2011 19.23
A: FANN General and development discussion
Oggetto: Re: [Fann-general] size ANN training input



Normalization you mean?
Does it depends on the training algorithm? Having your data normalized is a
good practice.

On Nov 2, 2011 2:49 PM, "Andrea_Viano" <viano-***@public.gmane.org> wrote:

Hi,



Is there anyone that knows why training inputs of an ANN have to be less
than 1 to have a convergence? I tried with my regular input( 10000,or 8
etc), and I had a very high and constant error. If I divided training inputs
to reach numbers close to the unity I could find a solution of my ANN.



Can you help me to solve this problem?



Thanks



Andrea



Ing. Andrea Viano

Industry Division



_____

Logo_CFD_Engineering.jpg



CFD Engineering S.r.l.

Headquarters

Piazza della Vittoria, 7/2A - 16123 Genova (Italy)

Phone +39-010-8540050 <tel:%2B39-010-8540050> - Fax
+39-010-8631199 <tel:%2B39-010-8631199>




----------------------------------------------------------------------------
--
RSA&#174; Conference 2012
Save $700 by Nov 18
Register now&#33;
http://p.sf.net/sfu/rsa-sfdev2dev1

Continue reading on narkive:
Loading...