normalization for neural network

need clarification on normalization.

I have 3 inputs (with different units), 1 output and a total of 27 samples. 21 samples will be used for training and another 6 samples for testing. my inputs and output is not in range 0 to 1. so i need to normalize it.my questions is

Do i need to normalize it by column or row or input/output? let say, here are the first 3 samples of inputs and outputs

• inputs >> output
• 3 45 -10 >> 3
• 5 55 -3 >> 56
• 6 23 -4 >> 32

*note that all 3 inputs are in different units

so, do i need to normalize by row (3 45 -10) or by column (3 5 6) or using all inputs [3 45 -10; 5 55 -3; 6 23 -4] without considering the units at the same time?

i want to normalize in 0.1-0.9 range. can i used mapminmax?

Answer 1

[Greg Heath]

All of the current NN designs use mapminmax by default. However, for purposes of understanding, I use zscore (you can use mapstd) for pre-training analysis and plotting. For convenience, I let the design function use the default mapminmax just because it is a pain to remove it.

There is absolutely no reason whatsoever to use 0.1-0.9.

Why would you want to normalize columns???

Hope this helps.

Thank you for formally accepting my answer

Greg

[azie]

each column will has different dimensional units.so, should i normalize the input according to their units or i can just normalize according to input samples(which will have 3 different dimensional units)?

[Greg Heath]

To use the NNTBX, variables are rows, and samples are columns. Normalize each variable row according to its own summary stats.