I'm trying to use neural network partial least square proposed by Qin and McAvoy. The whole network trained based on the scaled values. I want to know how can I rewrite the general neural network equation (where parameters are trained based on scaled values) so that I can be able to find the prediction for original predictors. The transformation from normalized linear model to standard setting is straight forward, but how is it possible for non-linear models.

Any suggestion would be appreciated. Specially if there exist a function in R that can do the job. In general, is it suggested to scale inputs before using neural net or not?


migrated from stackoverflow.com May 14 '15 at 21:07

This question came from our site for professional and enthusiast programmers.

  • $\begingroup$ In general, one scales the ANN inputs to the range $[-1,1]$ or to have unit variance. Not sure about the rest, though. $\endgroup$ – Sycorax May 15 '15 at 2:22
  • $\begingroup$ Thanks for the comment. I found out even if I use the scaling, I need to apply the same scaling used on training for testing too. So I'm trying this to see how it works! $\endgroup$ – Mina.M May 15 '15 at 14:23
  • $\begingroup$ what do you mean by nonlinear model? You will apply your transformations only on features, and nonlinearities will happen in the network. You can hardcode weights so the network is doing scaling, however scaling the features directly will be much easier $\endgroup$ – rep_ho Oct 12 '17 at 7:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.