Quote:
Originally Posted by davidrcoelho
Hi everyone,
As I understood, noisy target is when we observe that for a same value of x we get different values of y.
Then, we model this function as a distribution P(yx) (instead of a deterministic one)
But later on in the book, Yaser says:
"This view suggests that a deterministic target function can be considered a special case of a noisy target, just with zero noise. Indeed, we can formally express any function f as a distribution P(yx) by choosing P(yx) to be zero for all y except y = f(x)".
My question is:
How we can consider a value of y is equal to f(x), since the function f is a distribution.
Suppose we have two different values of y for the same input x, what of these two y's is different to f(x)?

Just to clarify, the statement above from the book is talking about viewing a deterministic target as (a special case of) a noisy target, not the other way around as the title of this thread "Noisy Targets as deterministic target function" may suggest. A noisy target can be viewed as a deterministic target
plus noise, but not as a standalone deterministic target.
The case you mention, where there are two possible
's for the same
is a case of a noisy target, not a deterministic target.