I've been learning a little bit about computational learning theory, but most of what I've seen so far is related to supervised learning. Perhaps dimensionality reduction will be touched on, but not very in depth.

So I'm wondering - is there a formal notion of "generalization" in unsupervised learning that is analogous to the PAC model for supervised learning? I'm actually curious how we can get a sense of how "good" a particular unsupervised technique is without the notion of a loss function and labels.

  • 1
    $\begingroup$ Hi Marcel. There's also the site Artificial Intelligence SE where you may ask this question. $\endgroup$ – user34637 Apr 18 '20 at 6:16
  • 1
    $\begingroup$ @D.W. I don't fully understand the big problem behind this cross-posting and why people are discouraging it. People from different communities can provide different perspectives. This is a good advantage. Anyway, Stack Exchange should provide a more flexible way of moving questions. Currently, the system is very slow and inflexible, in my opinion. Btw, I am a moderator at AI SE. $\endgroup$ – user34637 Apr 19 '20 at 23:16
  • 1
    $\begingroup$ @nbro, discussing the merits of the policy is probably beyond the scope of this comment section. If you'd like to discuss, let's take it to a separate chatroom. You can find some of the arguments at the link I shared; and in any case, that link describes the current default Stack Exchange policy (if a site doesn't specify its own policy that overrides it), and it is policy here on this site that simultaneously cross-posting on another site is not allowed. $\endgroup$ – D.W. Apr 19 '20 at 23:19
  • 1
    $\begingroup$ @D.W. As far as I understand, an answer to SE question doesn't make something a rule. It only shows the consensus of the people. I've read some answers related to the topic in the past and they usually say "don't do it because don't do it". That argument is as poor as my will to read those answers again. $\endgroup$ – user34637 Apr 19 '20 at 23:21