The usual strategy in neural networks today is to use min-max scaling to scale the input feature vector from 0 to 1. I want to know if the same principle holds true if our inputs have a large dynamic range (for example, there may be some very large values and some very small values). Isn't it better to use logarithmic scaling in such cases?
$\begingroup$
$\endgroup$
If it is a classification problem, then you will use sigmoid or softmax to make the output value in (0,1) and all the value must sum to 1 as per the rule of probability.
-
$\begingroup$ I want to know about the scaling of inputs to the neural network. Rather than using (x-min)/(max-min), can we use logarithmic scaling $\endgroup$ – Rishabh Jain Jun 21 '20 at 13:31
-
$\begingroup$ If you want your model to learn properly you should use normalized data. There no gurantee other transformation won't work. It might. But normalising data is a standard practice. $\endgroup$ – SrJ Jun 21 '20 at 13:42