Agostino.Manzato@osmer.fvg.it
2004-Jun-16 15:03 UTC
[R] Is input pre-processing needed for nnet module?
Hi, I used the nnet R module to classify my data using Neural Networks: nnet(input_matrix, obs_vect, size=h, linout=FALSE, entropy=TRUE) I used as NN input my "raw" data. After that I tried to use the normalized input data (with z-scores, i.e. mean=0 and std=1) and have found NNs with a little smaller Cross Entropy Error. My question is: Is it *wrong* to feed nnet directly with the raw input data? I found in http://www.faqs.org/faqs/ai-faq/neural-nets/part2/section-16.html that it depends on the minimization training algorithm: - "Steepest descent is very sensitive to scaling. - Quasi-Newton and conjugate gradient methods... therefore are scale sensitive. However,... are less scale sensitive than pure gradient descent. - Newton-Raphson and Gauss-Newton, if implemented correctly, are theoretically invariant under scale changes..." I know that nnet is a Quasi-Newton algorithm, so it make sense that I found a small improvement using the normalized data. Can someone confirme if it is really so? Thank you very much! -- Agostino.Manzato at osmer.fvg.it
Prof Brian Ripley
2004-Jun-16 15:53 UTC
[R] Is input pre-processing needed for nnet module?
If you read the reference on the help page, you will find out the answer. After all, as the DESCRIPTION says, this is support software for a book. On Wed, 16 Jun 2004, Agostino.Manzato at osmer.fvg.it wrote:> Hi, > I used the nnet R module to classify my data using Neural Networks: > nnet(input_matrix, obs_vect, size=h, linout=FALSE, entropy=TRUE) > I used as NN input my "raw" data. > After that I tried to use the normalized input data (with z-scores, > i.e. mean=0 and std=1) and have found NNs with a little smaller > Cross Entropy Error. > My question is: > Is it *wrong* to feed nnet directly with the raw input data? > > I found in > http://www.faqs.org/faqs/ai-faq/neural-nets/part2/section-16.html > that it depends on the minimization training algorithm: > - "Steepest descent is very sensitive to scaling. > - Quasi-Newton and conjugate gradient methods... therefore are scale > sensitive. However,... are less scale sensitive than pure gradient > descent. > - Newton-Raphson and Gauss-Newton, if implemented correctly, are > theoretically invariant under scale changes..." > > I know that nnet is a Quasi-Newton algorithm, so it make sense > that I found a small improvement using the normalized data. > Can someone confirme if it is really so?-- Brian D. Ripley, ripley at stats.ox.ac.uk Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/ University of Oxford, Tel: +44 1865 272861 (self) 1 South Parks Road, +44 1865 272866 (PA) Oxford OX1 3TG, UK Fax: +44 1865 272595