I've implemented a MLP NN and I would like to get a clarification on a question where I haven't been able to find answer in books.
I've used couple of different normalisation techniques for my inputs and trained the network with all of them. Surprisingly, when I used minmax scaling which ranges the inputs between 0 and 1. This affects regularisation as well, for example if I train a network with the same data normalised with different technique such as Gaussian, I normally need weight decay to make good predictions on similar dataset, but scaling inputs to 0-1 seems sufficient to make better generalisation on similar dataset without any weight decay. Is it possible that when inputs are scaled 0-1 it results in lower weights which makes generalisation better?