Open utterances-bot opened 2 years ago
So this is a lot more concrete ML, right? Using the dreaming images as proxies for what the model knows but might deform?
Yes, it's pretty specific to scenarios which include an evaluator/reward model trained via gradient descent to e.g. approximate human feedback. I'll queue up a question on other possible applications.
Contrastive Dreaming
Description placeholder
https://paulbricman.com/hypothesis-subspace/?stackedPages=%2Fcontrastive-dreaming