4
$\begingroup$

We know that the theory of PAC-learning is distribution-free, i.e. assuming that the test and train distributions are the same, we have guarantees on learning the hypothesis.

Question: what if the train/test(evaluation) distributions don't match; is there any result that tells us we can still learn something (say when the training distribution is log-concave and the test distribution is uniform).

$\endgroup$
3
$\begingroup$

In general, the results are pretty strongly negative --- fairly strong assumptions are needed for something like this to work. As an extreme case, suppose that training and testing distributions have disjoint supports, so that the training sample can be nearly uninformative regarding the test performance.

That having been said, there is a body of research on learning (with) changing target concepts/distributions. Here are four fairly representative samples; once you do a follow-up search of who cites them and who they cite, you'll pretty much span the space:

https://link.springer.com/chapter/10.1007/978-3-642-34106-9_13

https://link.springer.com/article/10.1023/A:1007604202679

https://dl.acm.org/citation.cfm?id=269812

https://link.springer.com/article/10.1023/A:1007666507971

Also potentially relevant: Transfer learning https://en.wikipedia.org/wiki/Transfer_learning

$\endgroup$

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.