Neural Mapminmax option for a set

When I use mapminmax like,

net.inputs{1}.processFcns = {'removeconstantrows','mapminmax'};
net.outputs{2}.processFcns = {'removeconstantrows','mapminmax'};

it will mapminmax the whole set of:

net.divideParam.trainRatio = 70/100;
net.divideParam.valRatio = 15/100;
net.divideParam.testRatio = 15/100;

???

Or it would individually mapminmax values of valRatio and TestRatio based on the trainRatio minmax procedure ?

If its a common normalization, would´t the testset or the valset mess with the normalization of the trainset ? Andre

Answer 1

[Greg Heath]

If the total data set is assumed to be a random sample from the population and the trn/val/tst subsets are random samples from the total, then the expectation is that there is no way to tell that the val and tst subsets are not random samples from the population.

However, this is statistics. Therefore there is no guaranty.

As a result, it is not unwise to be aware of the summary statistics of those sets of data and any new data.

Hope this helps.

Greg

[Greg Heath]

The training record tr in [ net tr ] = train(net,x,t) contains the indices of the separate trn/val/tst subsets. You can always check to see if renormalizing the 3 subsets separately significantly affects the 3 error rates.