What is net.inputs​{1}.proces​sFcns doing to the inputs? How can a narxnet balance inputs equally?

Matlab documentation describes

net.inputs{1}.processFcns

as normalization step applied to both the input vectors and the target vectors and the default is set to mapminmax which normalizes inputs/targets to fall in the range [−1, 1].

If the default is used and have 3 inputs, is each input normalized individually? or all values combined? For example

I1: original range 0-0.05 --> Normalized so 0 is -1 and 0.05 is 1

I2: original range 0-0.15 --> Normalized so 0 is -1 and 0.15 is 1

I3: original range -2-3.5 --> Normalized so -2 is -1 and 3.5 is 1

OR everything is normalized to -1to1 so that

I1: --> Normalized so 0 is -0.23 and 0.05 is -0.25

I2: --> Normalized so 0 is -0.23 and 0.15 is -0.22

I3: --> Normalized so -2 is -1 and 3.5 is 1

I have trained a 2 input narxnet and a 5 input narxet, individually. The 2 input NN is stable but doesnt meet the test performance goal. When I include 3 more inputs, the 5 input narnext is stable and meets the test perfomance goal. However, the new 3 inputs, which have larger values, up to 70 times larger, takes over the prediction so the original 2 inputs are not used much. How can the 5 inputs be more balanced (used equally) or having the original 2 inputs taking over the prediction and having the extra 3 inputs just to help meet the performance goal?