How to minimaze the number of neural network input?

Hi everybody, I am trying to choose which are the best input parameters for my feedforward neural network. Actually, I have 14 input parameters that give me good results in performance. However I've noticed that by using only 7 of them (choosen on the basis of my intuition), the performance increases. Is there therefore a way to know which are the best input parameters to increase the performance (probably reducing also the number of input parameters)? How to do it? Thanks

Answer 1

[Greg Heath]

Not enough details.

Finding the optimal combination is, in general, complicated. However, using stepwise or stepwisefit on a model that is linear in the variables generally yields a decent approximation.

In addition, if you start with a model that is quadratic in variables with squares and crossproducts but LINEAR IN COEFFICIENTS, the stepwise functions can still be used.

Otherwise, use a backward search with the NN inputs. For example

 1. Train with all
 2. Obtain the responses when each input is separately randomized. 
 3. Remove the variable who randomization degrades the performance the least 
 4. Repeat 2&3 until you can't stand further decreases in performance.

Hope this helps.

Thank you for formally accepting my answer

Greg

Answer 2

[Gino Massafra]

That's it. Thank you Greg