a

(a) In PAC learning, suppose we let the learner choose the feature points he/she would like classified instead of providing randomly drawn examples.

Suggest a way in which you might measure how much “learning” has taken place in this case, that is, what sort of performance criterion might you use to measure success after the learner has asked for some number of examples?

(b) For your answer to part (a), does this make learning easier or harder than standard PAC learning?