Steven Hansen

@Zergylord

Research Scientist at DeepMind. Slowly learning to learn fast. All opinions my own?

London, England
Vrijeme pridruživanja: lipanj 2009.

Medijski sadržaj

  1. 29. pro 2019.

    Aristotelian physics may be flawed, but it has been unfairly discarded in favor of Newtonian physics, despite the latter not solving all of the open problems in the field. Clearly the way forward is a hybrid system, wherein aether obeys the conservation of momentum.

  2. 19. pro 2019.

    Conference-decision-anticipation at an all time high. I need that sweet sweet email notification!

  3. 13. pro 2019.

    talk at I managed to ask a question afterwards: Q: Did you ever implement, or even attempt to implement, the "special case" where adversarial training is used to mimic a training dataset A: No. Hopefully this helps clear up some misconceptions ;)

  4. 13. pro 2019.

    finally released a paper on their DOTA AI. I've given them a lot of crap over the years for not sharing details in a timely manner (and I still stand by most of it that), but this sort of openness is still greatly appreciated.

  5. 11. pro 2019.

    Is it too late to rebrand entropy regularization as "the free-will prior"?

  6. 11. pro 2019.

    "now let's talk about consciousness" Can we not?

  7. 9. pro 2019.

    I like London, but it'll be so nice to just forget about Brexit and think about nothing but AI for a week. Day 1:

  8. 9. pro 2019.

    Great tutorial at Turns out that "dataset shit" is really hard.

  9. 5. pro 2019.
    Odgovor korisniku/ci

    Couldn't help thinking of this :)

  10. 28. stu 2019.

    What's that? An excuse for posting about my beloved baby bear? Don't mind if I do! Bear on chair:

  11. 28. stu 2019.

    How is there not a paper on the DOTA AI yet? I remember asking about this two years ago (link to exact timestamp), and he claimed there would be at least an arXiv paper after 5v5 results (which they had in 2018 and concluded this April).

    Prikaži ovu nit
  12. 7. stu 2019.
    Odgovor korisnicima
  13. 6. stu 2019.
    Odgovor korisniku/ci

    Oh wow, there truly is an XKCD for everything! I had this in mind:

  14. 5. stu 2019.
  15. 5. stu 2019.

    reviews are out! Good luck everyone! May your rebuttals fall on receptive ears, and your additional experiments be non-existent.

  16. 30. lis 2019.

    Feverishly working on preparing the tasks for an external just in time for . We hope these tasks represent an interesting challenge for the deep RL community. Excited to see what y'all can do with them! n/n back to work time

    Prikaži ovu nit
  17. 30. lis 2019.

    Memory Recall Agent! A new agent that combines 1) an external memory 2) contrastive auxiliary loss 3) jumpy-backpropagation for credit assignment Importantly, all of these pieces were validated through over 10 ablations! 5/n

    Prikaži ovu nit
  18. 30. lis 2019.

    Tasks! In addition to a standard train/test split based on partitioning some variable (e.g. color), we also pick a scalar variable (e.g. size of room). We can thus train on some values and test on unseen values inside the range (interp) or outside of the range (extrap) 4/n

    Prikaži ovu nit
  19. 30. lis 2019.

    Results! 1) Some of these tasks are hard! Underfitting is still an issue in RL 2) Extrapolation isn't impossible for Deep RL agents, but it requires the right inductive biases and is far from solved 3) Adding a contrastive loss to an external memory is a good thing to do 3/n

    Prikaži ovu nit
  20. 30. lis 2019.

    Excited to announce our work on memory generalization in Deep RL is out now! We created a suite of 13 tasks with variants to test interpolation and extrapolation. Our new MRA agent out-performs baselines, but these tasks remain an open challenge. 1/n

    Prikaži ovu nit

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:

    ·