4 ms·
It’s very tough to do research-level ML without the whole track. There are exceptions, but if you want to publish papers at DeepMind, you probably want the grad
by benreesman 3y ago
It’s very tough to do research-level ML without the whole track. There are exceptions, but if you want to publish papers at DeepMind, you probably want the graduate education.
But if you, like me, are happy to be an “understand and implement the paper” person instead of a “co-author the paper” person, that is eminently achievable via self-study and/or adjacent industrial experience. In fact, it’s never been easier as world-class education is more available on YouTube every day.
3Blue1Brown covers all the linear algebra, calculus, information theory, and basic analysis you need to get started. Ng and Karpathy have fantastic stuff. Hotz is writing a credible threat to PyTorch on stream/YT for the nuts and bolts accelerated computing stuff. Kilcher does accessible reviews of modern stuff. The fast.ai stuff is great.
This is all a lot easier if you can get a generalist/infrastructure role that’s ML-adjacent at a serious shop (that’s how I got exposed), but there’s enough momentum in open source now that even this isn’t a prerequisite.
I say fucking go for it, and if you want any recommendations on books or need someone to ping when you get stuck, feel free to email.
- tayo42 3y agoYeah im ok without doing real research, but just looking for a space with new interesting challenges. Ive been seeing how much is on youtube and stuff which made me start thinking this all might be more accessible then I thought initially. Thanks for the encouragement, yeah books would be interesting too.
- benreesman 3y agoOne metric for impact is how much damage it causes to established players. In the loop of voice -> agent -> voice, the cartel left only one edge out. Voice to text/latent? Whisper is good. LLMs? They’re everywhere. LLaMA 2 is practically state of the art and it’s licensed for commercial use. Latent/Text -> voice? Best you can get open is T5, which is ok, but it’s no Voicebox or whatever. They launch new SOTA’s every month or two, but no weights, no code. Stay dangerous. You want to work on that? I’ll do it with you. On books, to get up and rolling I’d recommend Raschka, Liu, and Mirjalili. The one even the PhDs should re-read is MacKay.
- yu3zhou4 3y agoDo you maybe recommend some open source projects to contribute to as a way of learning ML?
- benreesman 3y agoIf I had more time I’d be working on tinygrad every day.
- tayo42 3y agoInteresting, thanks for the recs. Will check them out!
- gvlachos 3y ago[dead]