scale data for NN

Hi,

How do I scale data in a neural network multilayer backpropagation?

Answer 1

[Greg Heath]

You do not have to scale the data because variables are AUTOMATICALLY scaled to {-1,1} with MAPMINMAX by NEWFF, FITNET, PATTERNNET AND FEEDFORWARDNET.

The command

type feedforwardnet

yields these commands

========================

% Inputs

net.numInputs = 1;

net.inputConnect(1,1) = true;

net.inputs{1}.processFcns = {'removeconstantrows','mapminmax'};

% Outputs

net.outputConnect(Nl) = true;

net.outputs{Nl}.processFcns = {'removeconstantrows','mapminmax'};

===================================================

If you wish, you can replace either occurrance of MAPMINMAX with 'mapstd' (zero-mean/unit-variance) or 'none'.

I prefer to use MAPSTD before creating the net to try to understand the input/output relationships via plots, correlation coefficients and outliers.

Then I accept the automatic minmax normalization instead of removing or changing it.

Hope this helps.

Greg

[mustafa]

Thanks Greg for ur reply. I want to build my program by own ( I dont use the toolbox). R there formula for scale?. Actually, i found this formula "I = Imin + (Imax-Imin)*(X-Dmin)/(Dmax-Dmin)" by searching in google but I dont know the reference it.

[mustafa]

so I see old ur comment . You used this formula "xn = -1+ 2*(x-xmin)/(xmax-xmin) ;". is this formula for scalling? if yes, can I know the refrence? So what do you meant the numbers(1 &2) in formula.

Answer 2

[Greg Heath]

1. Derive a linear tranformation xn(x) = a.*x + b such that

 xn( x = min(x) )  =  -1 

 xn( x = max(x) )  =  +1

2. Derive a linear tranformation xn(x) = a.*x + b such that

 mean(xn)  =  0

 var( xn ) =  1

Hope this helps.

Greg

[mustafa]

sorry Greg, it is not clear. plz I need some details about this.