Grzegorz Kossakowski

@gkossakowski

Proponent of dense representations. Previously: , hobo, at .

San Francisco, CA   Munich, DE
Vrijeme pridruživanja: ožujak 2009.

Tweetovi

Blokirali ste korisnika/cu @gkossakowski

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

  1. 23. sij

    Inflight entertainment

    Poništi
  2. 23. sij

    SFO ✈️ WAW

    Poništi
  3. 23. sij

    You should try 's training loop even if your model happens to fall into one of the broad categories they support like vision, nlp or tabular data. End of my today's praise fo . 4/4

    Prikaži ovu nit
    Poništi
  4. 23. sij

    Spot checking the examples returned by `most_confused` and `top_losses` from the Interpretation object revealed mislabelled training examples. These are little touches that are really handy in practice. 3/4

    Prikaži ovu nit
    Poništi
  5. 23. sij

    The big surprise comes from the fact that I was using exactly the same optimizer and learning rate as in the home-grown training loop. I also disabled weight decay and still 's training loop was producing a better model. Not sure why. 2/4

    Prikaži ovu nit
    Poništi
  6. 23. sij

    I've experimented swapping a home-grown PyTorch training loop with one and kept the embeddings classification model intact. To my surprise, classifier's performance jumped the most since this model was put together originally 1/4

    Prikaži ovu nit
    Poništi
  7. proslijedio/la je Tweet
    20. sij

    “In 1990 a generation of baby-boomers, with a median age of 35, owned a third of America’s real estate by value. In 2019 a similarly sized cohort of millennials, aged 31, owned just 4%”

    Poništi
  8. 20. sij

    The modern Stochastic Gradient Descent is marvelous. I've seen enough examples of its effectiveness and still can't wrap my head around how well it works.

    Prikaži ovu nit
    Poništi
  9. 20. sij

    I heard a while back from that deep learning models can be trained on noisy data. It's biased data that's problematic. I trained recently a graph embedding model with 50% of training examples being bogus and the model still achieved a remarkable performance

    Prikaži ovu nit
    Poništi
  10. proslijedio/la je Tweet
    16. sij

    The guy who invented the locomotive had to do an experiment to prove to himself that a vehicle could move itself simply by turning its own wheels 🚂

    Prikaži ovu nit
    Poništi
  11. proslijedio/la je Tweet
    15. sij

    The feedback will continue until morale improves. ✍️

    Poništi
  12. 15. sij

    That essay calls ML an "enabling layer" and that's the best term I have seen. Interestingly, Jeff Bezos calls the Internet an enabling layer in his 2007 TED talk: It's an entertaining and enlightening talk from the time TED represented quality

    Poništi
  13. 15. sij

    That essay was one of sources for my go-to rule that any startup with .ai in its domain name represents either a purposeful acquisition target by triggering the right keyword search or a lousy product thinking

    Poništi
  14. 15. sij

    The essay was published in 2018 but I stumbled upon it in 2019.

    Prikaži ovu nit
    Poništi
  15. 15. sij

    Looking back at 2019 reads, I find Benedict Evans' "Ways to think about machine learning" to be one of the most interesting frameworks of thinking of ML I reread it today and realized that it ages really way

    Prikaži ovu nit
    Poništi
  16. proslijedio/la je Tweet
    11. sij

    the programmer equivalent of a lawyer to represent my interests within various software systems

    Prikaži ovu nit
    Poništi
  17. 7. sij

    WAW ✈️👉 SFO

    Poništi
  18. proslijedio/la je Tweet
    7. sij

    this is absolutely nuts. The AI GPT-2 has learned to play chess moderately well (able to give bad human amateurs a game) – despite only being a text AI, learning from a corpus of chess notation text, and not having any concept of what a chessboard is

    Prikaži ovu nit
    Poništi
  19. That paper claims that MC Dropout is indeed used against adversarial examples. Phew, i wasn’t off with my intuition. And as it usually goes with good ideas for pressing problems: someone is already on it.

    Prikaži ovu nit
    Poništi
  20. Keeping Dropout on is called Monte Carlo Dropput and can be thought as having an ensemble of networks instead of a single network. I wondered if MC Dropout could be used to counter adversarial examples that are tuned for particular weights and here it is:

    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:

    ·