Does increasing noise in the data would help to achieve better learning ability by the network, doesn't make any difference or it depends on the problem being solved? How does it affect the generalisation process in overall?