Alice's Adventures in a Differentiable Wonderland

(arxiv.org)

137 points | by henning 3 days ago ago

25 comments

  • bwfan123 9 hours ago

    this 3 page classic [1] captures most of the core ideas and explains it in a manner anyone with basic calculus background can understand - "Learning representations by back-propagating errors"

    [1] https://gwern.net/doc/ai/nn/1986-rumelhart-2.pdf

    • pstuart an hour ago

      I took calculus over 30 years ago and have never really used it -- I'll put your conjecture to test (sample size: 1). Will let you know if your conjecture is true ;-).

  • 0cf8612b2e1e 10 hours ago

      The corresponding row vector is denoted by x^T when we need to distinguish them. We can also ignore the transpose for readability, if the shape is clear from context.
    
    I am tilting at windmills, but I am continually annoyed at the sloppiness of mathematicians in writing. Fine, you don’t like verbosity, but for didactic purposes, please do not assume the reader is equipped to know that variable x actually implies variable y.

    All that being said, the writing style from the first chapter is very encouraging at how approachable this will be.

    • runeblaze 6 hours ago

      It is weird to be honest. I first learned Coq and then started taking upper level maths classes. My group theory proofs were panned by my TAs as overly verbose, very precise, and I was specializing on H_1 and H_2s everywhere and having IHns flying around like crazy because I could not fathom how one proves things without formally connecting things up.

      Then my profs told me I was not “wrong”, but proofs or expositions are to most mathematicians not programs (ha! How did I not know. You teach me natural deduction and expect me not to program?), more like convincing arguments/prose. At some point one abstracts.

    • JadeNB 7 hours ago

      > I am tilting at windmills, but I am continually annoyed at the sloppiness of mathematicians in writing. Fine, you don’t like verbosity, but for didactic purposes, please do not assume the reader is equipped to know that variable x actually implies variable y.

      I am a practicing mathematician who felt the same way you did when I started, and who still writes their papers in a way that many of my colleagues feel is gallingly pedantic. With that as my credentials, I hope I may say that it can be much worse as a reader to read something where every detail is spelled out, because a bit of syntactic sugar begins to seem as important as the heart of an argument. Where the dividing line is between precision and obfuscation depends on the reader, and so inevitably will leave some readers on the wrong side, but a trade-off does have to be made somewhere.

      • prisenco 3 hours ago

        Could there be a compromise where the verbosity is kept but the key points are highlighted, grouped or presented in a different color.

        I would certainly appreciate if math papers were more explicit and "hand-holding" but understand why trained mathematicians would find that tedious.

      • zmgsabst an hour ago

        I wish the formality would be included in an appendix — as someone who has had to implement a lot of things (and more than once, found errors).

        But I agree with your general point: understanding the recipe and general thrust of the approach is often more important, because even if the exact proof misses some technical detail, that can often be patched.

  • superjose 13 hours ago

    Wow, kudos to the Author. Very easy to digest, beautifully crafted, and took the time to explain the concepts when most places take them for granted.

    • canyp 4 hours ago

      Well, kudos to your one-line comment too because now I am encouraged to actually read this.

  • magnio 13 hours ago

    This looks like a good practical companion for a more theoretical text, such as Deep Learning by Bishop.

  • canyp 4 hours ago

    Beautifully formatted and has the right combination of code and theory for noobs like me. Strong vibes for Simone right now, hero of the people.

  • odyssey7 12 hours ago

    It would be nice if arXiv included a small-layout pdf or native epub option for e-readers. Now that they serve the Tex files and are experimenting with HTML, it feels like a natural step.

  • dunefox 6 hours ago

    And I just bought the physical book...

    • canyp 4 hours ago

      I do that all the time to support authors, plus the physicality of a tangible book is irreplaceable. In fact, I did that just today with a different book.

  • fossa1 14 hours ago

    Glad to see JAX featured alongside PyTorch. JAX still feels like the best-kept secret in deep learning

  • _giorgio_ 13 hours ago
  • kittikitti 13 hours ago

    Although I love this, it's not peer reviewed and I don't trust arxiv.

    • SiempreViernes 12 hours ago

      Actually, it is peer reviewed following the standard practice for books: some other people read it and provided feedback as evidenced by the Acknowledgments section.

    • odyssey7 12 hours ago

      It’s more a book than academic research.

      The funny thing about books is that authors in free societies are allowed to self-publish whatever they want. The norms are different and, frankly, more democratic and with less gatekeeping.

    • esafak 12 hours ago
    • ethan_smith 8 hours ago

      arXiv is a preprint server trusted by the scientific community for decades - papers there often undergo peer review later, and many top ML researchers publish their work there first for faster dissemination.

  • ProofHouse 13 hours ago

    Damn beeeeefffffyyyyy. Need the month to eat ten pages a day, Tnx looks awesome. Could append diffusion too ultimately