Neural Network cost function normalization

Hello,

I am using feedforward network in MATLAB.

My question is "what is the definite procedure of Normalization?".

Code: net.performParam.normalization='percent';

In the help, "'percent', which normalizes errors between -1 and 1" is written.

Then does it mean that normalization makes minimum value of errors to -1, and maximum value to 1?

For example, ground truth data is "1 2 3 4"

and the NN's output is "0.9 1.95 3 4.1"

then the errors would be "-0.1, -0.05, 0, 0.1".

Then, after normalization, will it become "-1 -0.5 0 1"?

Thanks regard.

Answer 1

[Sahil Jain]

The documentation for "nnparam.normalization" says "If it is set to 'percent' then percentage errors are used, relative to the ranges of the original target data, and errors will be mapped to the range [-1,1]."

Therefore, if "t" and "y" are the ground truth and output predictions respectively, then the normalized error values would be

[Dong Youn Lee]

Thanks a lot!

[少快 王]

Why is the range of the error not max(e) - min(e), but max(t)-min(t)? I think en(error normalization) should be：en=2*(e-min(e)) / (max(e) - min(e))-1