Dimitry Tegunov @dtegunov.bsky.social · May 3 Instead, @martenchaillet.bsky.social developed an unsupervised training procedure where the model is asked to score a reconstruction with good alignment lower than one with bad alignment. No pre-defined values – just a contrastive loss with triplets of examples.
9 likes 1 replies
?
Replies
But we're starting with bad alignments! Where do we get the good ones? We don't. We just make the existing alignments even worse by adding different types of noise to the parameters. You know: to outrun a bear, run faster than the other person.