0
$\begingroup$

I'm wondering what the difference between conditional learning and supervised learning is - especially in diffusion models? Am I correct to assume that diffusion models are supervised because in training, each noisy image is "labeled" with the original one but conditional learning would mean that in training, further labels (such as text data) would be included? And what do "self-supervised" and "semi-supervised" refer to in contrast?

And is it possible to do inference on a conditionally trained model with only lets say an image and no labels?

$\endgroup$
3
  • $\begingroup$ well one learns $p(y|x)$ and the other one learns $p(x|y)$ $\endgroup$
    – Alberto
    Apr 11 at 12:34
  • $\begingroup$ and you can always swap them with the Bayes rule, though it might be intractable $\endgroup$
    – Alberto
    Apr 11 at 12:34
  • $\begingroup$ Yes that is pretty much correct. Semi-supervised is an instance where you may have labels on for some of the data points (e.g. in node classification, you may only have labels for a subset of the nodes). Self-supervised is where you don't have any labels for your data. I'd say that self-supervised you use your data to learn useful representations, whereas unsupervised learning (where you also don't have labels) is more about looking for patterns and making predictions based on these patterns, such as clustering. $\endgroup$
    – David
    Apr 11 at 12:36

0

You must log in to answer this question.