Arseny Khakhalin

@ampanmdagaba

Assistant prof (Neuro) at . Networks, neurons, data, AI, complexity, beauty, platonic forms.

Vrijeme pridruživanja: lipanj 2012.

Medijski sadržaj

  1. prije 3 sata
    Odgovor korisnicima

    Yes, but not quite? If you ever tried to keep a knowledge base (I guess most of us did?), it's hard to do it in a "continuously updated review" mode! Too much ongoing rework. As a neuro person, I think of science as an axon growth cone: lots of probing, but limited consolidation.

  2. The most iconic video of all times (the very first viral video ever produced!): now stabilized, super-resolutioned, and re-colorized. I feel like it's a milestone of sorts. The closest we ever got to a time machine. Unbelievable!

    Prikaži ovu nit
  3. 3. velj
    Odgovor korisnicima

    Oh!! So that's the name of the artist spiritually behind my favorite animated cartoon of all times!?! TIL! :))

  4. 30. sij
    Odgovor korisnicima

    Focus length! (And then there's also face asymmetry that Natee didn't mention, but for those of us with broken noses and crooked smiles, it can make photos almost unrecognizable, compared to the mirror :)

  5. 30. sij
    Odgovor korisniku/ci

    Got inspired by you! import matplotlib.pyplot as plt f = lambda x,k,a: (x if k==0 else f(a*x*(1-x),k-1,a)) y = [f(.21,17 + i % 19, 2.5+1.5*i/10000) for i in range(10000)] plt.plot(y,'.',markersize=1); It's prob further compressable, but I can't do it :)

  6. 28. sij

    Or look at this one. A model that can recognize characters (Cyrillic! Must be Soviet), or read Russian (Glory to Soviet Builders), AND access to texts (Soviet banners are always red) would have colored banners red. A picture-only model cannot do it. A test for _Understanding_!

    Prikaži ovu nit
  7. 28. sij

    What would it take for it to connect the dots? (Another photo from my archive that no model have seen before). Recognize tokens (a portrait of a tyrant), and objects (flags) - jump to an easy conclusion. But it's a great test, as color photos with Stalin almost don't exist.

    Prikaži ovu nit
  8. 28. sij

    But the model doesn't know that! And we can see the same limitations here, in spirit, that GPT-2 demonstrated with language. Both models can do correlations, like coloring a single flag, or a well-known situation. It is "aware" of Soviets. But it cannot connect the dots.

    Prikaži ovu nit
  9. 28. sij

    by will evolve into a powerful test of general intelligence, maybe a better one than text-generation. Below is a photo from my family album, colorized by 's free deoldify ( ) But something is amiss here. Do you see it? (..thread)

    Prikaži ovu nit
  10. 27. sij

    TFW you think you are kinda familiar with a field, but then discover a classic study you never heard about.

  11. 26. sij
    Odgovor korisniku/ci

    I keep obsessing about this topic, and just looked whether it would color grandiose Red Square 1st of May parades red. You know, it kinda does! Hesitantly, again (not all flags are red, and it failed to grasp that balloons should be red too), but it "knows" about the Soviets!

  12. 26. sij
    Odgovor korisnicima

    I also learned that it doesn't know much about fire :) Catastrophic files, at least. Probably with them being such a rare topic. Looks surreal, doesn't it?

  13. 26. sij
    Odgovor korisnicima

    I played with the free version for quite some time today. For example, below it seems to, very hesitantly, notice Soviet symbols and correctly guess red. By the way, is the color intensity a function of model's certainty about color assignment? Or only borders? or both?

  14. 26. sij
    Odgovor korisniku/ci

    Haha, I'm glad to have found a missing cultural competency! :) It's fun! But in some way, it's almost generalized intelligence kind of bar. Like: old photo, Cyrillic letters (or maybe, one day, even a translation: "Glory to the USSR") - has to be red! But it's a very high bar!

  15. 25. sij
    Odgovor korisnicima

    While this Old Believers chapel from early 20 century is, I'd say, rather successful! It errs on the "old and brown" side, but still adds reasonable splashes of color here and there!

  16. 25. sij
    Odgovor korisnicima

    All these flags should (obviously! :) be red too! But the model is completely oblivious of that :) Isn't it funny? It's like cultural bias; something, potentially, to be studied!

  17. 25. sij
    Odgovor korisniku/ci

    I have an update on our old discussion: you know what obvious thing is really underrepresented in the training set for (at least DeepAI's) deoldify? Soviet photos! :) Not surprisingly. But as a result, it doesn't know about Soviet flags. Like, all of this should be red :)

  18. 24. sij
  19. 24. sij
    Odgovor korisniku/ci

    Gosh, I thought you are messing with us! But Google is on your side! How is it even possible?? Like, who would win: a lofty spacious precisely manufactured 14 ml tube with lots of fluid; this flat pewter boi?..

  20. 21. sij
    Odgovor korisnicima

    One could argue that Prokudin-Gorsky "invented" it, except that he prob hated this "invention" as for him it defied the very purpose of his work! So from this point of view he definitely had not invented it :) Like look at the smoke on this photo:

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:

    ·