training set bounds

Question 1

why is the question, Is it possible to derive a similar learning guarantee if S is the training sample?

Answer 1

might not have enough labelled data to afford an independent test set
gain more insight into learning
inspire new algorithms

Question 2

what do we do to prepare for Occam’s razor bound

Answer 2

we take and adapt a bayesian idea
before seeing training set S
we define a prior probability distribution over classifiers , p(c)
-> represents ‘‘our’ bet on the candidate classifiers