noahtren

@noahtren

AI, tools for thought, cog sci, etc. | DMs are open :)

Vrijeme pridruživanja: travanj 2019.

Tweetovi

Blokirali ste korisnika/cu @noahtren

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @noahtren

  1. Prikvačeni tweet
    4. sij

    ok, last visualization for Twitter, this is the best one by far. check out the code here:

    Prikaži ovu nit
    Poništi
  2. 1. velj

    what are some good online courses that are very project-oriented? I learn best by doing, but I can't think of any accessible starter projects in what I want to learn (additive manufacturing and synthetic bio). I'm thinking something like the format of nand2tetris

    Poništi
  3. 31. sij

    Cooperative Communication Networks can learn robust *and* human-readable representations. They could be used to help visualize a variety of complex problems. Really exciting to be working on this!

    Poništi
  4. 27. sij
    Poništi
  5. 23. sij

    just got an assignment to write about Turing's imitation game. thinking about using Write With Transformer from to help me write it, but not sure if it would help or hurt my argument 😛

    Prikaži ovu nit
    Poništi
  6. proslijedio/la je Tweet
    22. sij

    This network has never been exposed to the letters L, A, D, or P. And yet it has discovered their shapes! Communicating efficiently over a noisy channel forces these networks to converge to noise-robust representations, similar to how we humans created our own alphabet

    Prikaži ovu nit
    Poništi
  7. 19. sij

    I started thinking about this specifically because some operations are commutative/symmetric while others aren't, but there's no way to know this besides learning things like "commutative" and "associative" properties

    Prikaži ovu nit
    Poništi
  8. 19. sij

    what do you think the chances are that some future math notation (in <50 years) will make our current notation look like Roman Numerals? that is, it will make many operations much easier and more intuitive

    Prikaži ovu nit
    Poništi
  9. 18. sij

    nocode is cool, but who's making an interpreter that translates nocode to nobytecode to run on a noVM? or better yet compiles directly to noassembly? (serious joke question -- can systems programming be done visually and if so what might it look like?)

    Poništi
  10. proslijedio/la je Tweet
    5. sij

    Dimensions of Dialogue project expanded upon by ! Added new features inspired by POET and a new TF2 library.

    Poništi
  11. 4. sij

    so fun seeing that the final result can vary greatly even with the same starting parameters. makes me want to get a dozen instances running in the cloud

    Prikaži ovu nit
    Poništi
  12. 3. sij

    first timelapse of evolving a visual language. interesting to see shapes go quickly from square grids to more organic patterns, almost like ink on a page. 🤔 will run again with some different parameters, hopefully to reduce checkerboard artifacts and get more distinct glyphs

    Prikaži ovu nit
    Poništi
  13. proslijedio/la je Tweet
    30. pro 2019.

    Names are hard so calling this one "JS Bubbles". Also using colors to show the relationship between the bubbles and the symbols instead of arrows. This makes it easier to move the bubbles around. Pretty happy with the results so far 🤞

    Poništi
  14. 2. sij

    trying to use automatic domain randomization, where the communication channel gets increasingly noisy as the symbols get more universally "readable" (at least by a CNN, maybe not ppl 😋)

    Prikaži ovu nit
    Poništi
  15. 2. sij

    the end goal of the experiment is an extremely readable alphabet, brought to you by deep learning™. readable meaning that if you squint your eyes, tilt your head, etc., it is easier to read than English or other natural languages

    Prikaži ovu nit
    Poništi
  16. 2. sij

    evolving an artificial alphabet that is robust to noise. model is still converging, but this snapshot is super fun. looks like ancient runes, or an alien language 👽

    Prikaži ovu nit
    Poništi
  17. 30. pro 2019.

    interestingly, the generated images have different underlying patterns after each run. 🤔 this is probably because the task is too easy (no noise + few unique symbols) and the neural networks quickly converge to a good solution, depending on however they were initialized

    Prikaži ovu nit
    Poništi
  18. 30. pro 2019.

    what's happening here? two neural networks are being trained to work together to communicate information in the form of an image. one generate images, and the other predicts what an image means. in this simple case, an image represents a number 1-26 (or letter a-z)

    Prikaži ovu nit
    Poništi
  19. 30. pro 2019.

    implementing 's Dimension of Dialogue work in TensorFlow 2 and getting some good results after just a few hours of coding 😀 I think the real fun will begin once some noise is introduced to the communication channel. here's an example of some early gray-scale images:

    Prikaži ovu nit
    Poništi
  20. proslijedio/la je Tweet
    28. kol 2019.

    New work in my Dimension of Dialogue series :) Two neural nets learn to communicate through their own emergent visual language. The resulting alphabet is a product of their adversarial and cooperative relationship. Here set in clay tablets.

    Prikaži ovu nit
    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·