from Hacker News

Ask HN: Self-directed learning path for machine learning in 2023?

by optbuild on 6/25/23, 5:58 AM with 5 comments

I want to learn machine learning, but, not in a hand wavy way. I want to go deep to understand and appreciate major works of the past, the present and future. Only way to do that is to understand the research papers and implement them on real datasets. For that I have to understand the math behind it.

I have calculus and some matrix algebra (not proof based linear algebra) background and I can program well.

Is there any guide equivalent to teachyourself CS but for machine learning and deep learning?

If not, then can you suggest open courses (preferably with assignments) or books from where I can learn both the required math and machine learning?

  • by mindcrime on 6/25/23, 6:23 AM

    If you're content with the current orthodox view / approach(es) then just browse around Github and find one of those "Awesome X" lists like "Awesome Machine Learning"[1], "Awesome Deep Learning"[2], "Awesome Artificial Intelligence"[3] and so on, and go to town.

    If you want to go deeper, including taking a step back in time and retracing the path(s) taken, to explore whether or not you might want to choose a different fork... well, that's doable, but it's a slog. I should probably write up a reading list for this approach and put it up on Github. It leads to some weird places though... like right before jumping over to HN and noticing this post, I'd been spending the last hour or so trying to track down two obscure Russian books on neural nets from back in the 1980's / 1990's... where print copies do not appear to be available in the US (and definitely not translated to English) and the nearest library to me with a copy of the one is the Library of Congress in D.C.

    Do you need to go down that particular rabbit-hole? Probably not. I'm just particularly interested in revisiting some earlier techniques / theorizing about NN's, that have fallen out of favor and aren't really taught much anymore.

    [1]: https://github.com/josephmisiti/awesome-machine-learning

    [2]: https://github.com/ChristosChristofidis/awesome-deep-learnin...

    [3]: https://github.com/owainlewis/awesome-artificial-intelligenc...

  • by discordance on 6/25/23, 3:18 PM

    I enjoyed the fast AI course: https://course.fast.ai/

    and Karpathy's course: https://karpathy.ai/zero-to-hero.html

  • by Greamy on 6/26/23, 9:13 AM

    If you really want to take the whole path for the sake of it, a quite agreed-upon path would look something like this (it's been referred to it a lot on Reddit, I think it's called the harsh guide to ML or something): -Take the Elements of Statistical Learning by Hastie and Tibshirani; Really great textbook that has all the in-depth mathematics for all the classical ML you need plus exercises - you can get Andrew Ng's Coursera course, it's still really good and relevant, and it takes you through all the ML you need plus exercises in both Python and R. - Go through the Deep Learning book, it goes through the details of DL, why and why not it is related to ML. - After that you're more or less in open waters, you have all the background, and you're left to figure out whatever you want for yourself. To get more into newer research in computer vision, just search ArXiv for all the still relevant papers and try to o through them and understand them. Trying to implement them yourself from an official Git repo helps tremendously, so definitely do that. Also if you're more into classical ML, go on keggle and try some of the contests out and see if you can manage to do anything. At the beginning, you won't, but read through the best solutions and see what they've done.

    Good luck with the whole process! I believe it is quite challenging to learn it all from the bare metal maths but definitely worth it and quite rewarding.

  • by randomcatuser on 6/25/23, 6:09 AM

    nice! i'm actually building an app that will map&distill knowledge (with GPT) to create guided courses in domains like this. basically transform long-form courses/articles/textbooks into duolingo+messenger.

    wanna help?