AdVAN

@advadnoun

machine nature

Vrijeme pridruživanja: veljača 2018.

Tweetovi

Blokirali ste korisnika/cu @advadnoun

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

  1. 6. velj

    I made a site! Here's a little post on cooperative communication networks

    Poništi
  2. 5. velj
    Prikaži ovu nit
    Poništi
  3. 5. velj
    Prikaži ovu nit
    Poništi
  4. 4. velj
    Poništi
  5. 4. velj
    Poništi
  6. 4. velj

    Examples of the output of agents that have essentially solved sending 10,000 classes with 64*64 pixels each against an agent adding noise. 10,000 classes with 4096 pixels is nothing.

    Poništi
  7. 4. velj

    Oh and # of classes, of course.

    Prikaži ovu nit
    Poništi
  8. 4. velj

    I need to do a curve/heatmap of various levels of noise compared to various image resolutions for communicating images. NN parameters may be a factor, but I figure the image dimensions are the bottleneck

    Prikaži ovu nit
    Poništi
  9. 4. velj

    And it isn't *just* initialization: if you've been following this project, you know that when noise is added instead of interference from an adversary, the images for each class closely converge between the 2 networks!

    Prikaži ovu nit
    Poništi
  10. 4. velj

    By way of example, you can probably tell which two of these are from one NN and which other two are from another.

    Prikaži ovu nit
    Poništi
  11. 4. velj

    I don't know if I mentioned this before, but when trained with an adversarial agent adding interference instead of random noise being used, two communicating neural networks will develop their own "accents."

    Prikaži ovu nit
    Poništi
  12. 4. velj

    Removing the embedding that projects from the number of categories to one image channel and instead relying only on AdaIN provides very straight-edged results!

    Prikaži ovu nit
    Poništi
  13. 4. velj

    Using adaptive instance normalization a la StyleGAN in a CCN with adversarial interference to generate larger images representing many more classes! Here're some cherry-picked, early outputs.

    Prikaži ovu nit
    Poništi
  14. 4. velj

    Interesting results with more pixels and more classes in this CCN...

    Poništi
  15. 1. velj

    Higher-res example of the CCN-CPPN

    Poništi
  16. 31. sij

    Reversible 8-bit lava lamp

    Poništi
  17. 31. sij

    What's really exciting is having an animation from the CCN-CPPN!

    Prikaži ovu nit
    Poništi
  18. 31. sij

    If anyone wants code for this process, I can make a lil' example colab notebook.

    Prikaži ovu nit
    Poništi
  19. 31. sij

    The rbf should have as many spaced-apart gaussians as you have categories, and now you've expanded your variable to be the same shape as if you had one-hot encoded it... but it will be as if you had some contribution from each embedding matrix that your number is near!

    Prikaži ovu nit
    Poništi
  20. 31. sij

    Here's the plan: you want to generate linearly spaced numbers from 0 to however many categories you have. Each one will become a frame in your animation. You then use a radial basis function to essentially slide gaussians over each linearly spaced number.

    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:

    ·