I need to use a Multi Layer Perceptron Network in order to perform some non-linear regression. Any ideas if it's possible to perform a task like that and how? Which activation function should be used for that purpose?
thanks!
-
$\begingroup$ I think this question is very simple and will probably be closed, but I'll reply anyway in case it is useful. $\endgroup$ – Trylks Oct 28 '13 at 10:22
-
1$\begingroup$ This question seems more suitable for Cross Validated. $\endgroup$ – Kaveh Nov 7 '13 at 5:24
-
$\begingroup$ What research have you done? We expect you to do some basic research on your own before asking. There is a lot written on multi layer perceptron networks, including on activation functions; perhaps you should start by reading some of that? $\endgroup$ – D.W. Dec 2 '13 at 0:37
$\begingroup$
$\endgroup$
Any normal activation function should do. Multi-layer perceptrons are universal approximators, so you should have no problems.
You can find this even on wikipedia. The limitation for perceptrons is linear separability. Intuitively, each new layer allows to classify inputs according to as many parameters as neurons in the the previous layer obtaining a classification in as many dimensions as neurons in the next layer. In short, for one layer you have to keep linear separability, but according to that linear separability you are kind of "bending" (non isomorphic transformation) that hyperspace into a different one, where linear separability may be possible, or at least one layer closer.
