Noisy targets:
Quote:
" Indeed, we can formally
express any function f as a distribution P (y I x) by choosing P (y I x) to be
zero for all y except y = f (x) . Therefore, there is no loss of generality if we
consider the target to be a distribution rather than a function"

I am not able to understand how is no loss in generality ensured by considering target distribution and not target function ?