11 Followers
37 Following
71 Posts
⟨ Machine Learning | Physics ⟩ and other stuff too
@pbloem Also, I'd love to hear more from the experts what their takes are on this paper. I think it will be talked about soonish in the Eleuther AI Diffusion Reading Group (https://github.com/tmabraham/diffusion_reading_group) .
GitHub - tmabraham/diffusion_reading_group: Diffusion Reading Group at EleutherAI

Diffusion Reading Group at EleutherAI. Contribute to tmabraham/diffusion_reading_group development by creating an account on GitHub.

GitHub

@pbloem It's a lovely paper! A take I've seen (and enjoy) is that perhaps the name "step-by-step-dedegradation" is a better name than "diffusion models."

I still really enjoy that there are so many insights of other fields that the usual "uncorrelated Gaussian pixel-wise noise" has connections with so many other areas of Maths & physics that allows one to, say, have a good starting guess of hyperparameters (the noise schedules!) but also you can forget all of that and it just... still works :)

@TedUnderwood Finally went around to reading this, good stuff!
@Eamon1916 So glad they skipped 3k
@lkngrrr @astatide kid named third party web scraper:
@nsaphra I've been on and off, but recently took a probability class with Machine Learning and it was very very pretty. I advocate this path

@chrisoffner3d There are some forks of Mastodon (and more specifically, its front end? I wanna say... but maybe it requires the backend to play nicely) that allow for search and/or search for user's posts.

But yeah, it's very frustating, even if Twitter search is also bad

@asayeed Kid named people without inner monologue:

@tess Reminds me to what happened to Infinity Train.

Digital Hoarders keep winning.

@manifest_irony @briankrebs only accepted in the US though