Ofir Nachum

@ofirnachum

Research Scientist at

Vrijeme pridruživanja: rujan 2012.
7 fotografija ili videozapisa Fotografije i videozapisi

Tweetovi

Blokirali ste korisnika/cu @ofirnachum

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

  1. 8. sij

    Many of our group's recent papers - DualDICE, ValueDICE, GenDICE, AlgaeDICE - can be framed as applications of this duality. Still, lots of potential remaining applications for others to explore, and lingering questions of how these formulations interplay with stoch. opt methods.

    Fenchel-Rockafellar duality is a powerful tool that more people should be aware of, especially for RL! Straightforward applications of it enable offpolicy evaluation, offpolicy policy gradient/imitation learning, among others
    4 proslijeđena tweeta 14 korisnika označava da im se sviđa
    Poništi
  2. 8. sij

    Fenchel-Rockafellar duality is a powerful tool that more people should be aware of, especially for RL! Straightforward applications of it enable offpolicy evaluation, offpolicy policy gradient/imitation learning, among others

    31 proslijeđeni tweet 108 korisnika označava da im se sviđa
    Poništi
  3. 17. pro 2019.

    Code is here: Hope this will serve as a strong baseline for future work to improve on!

    1 proslijeđeni tweet 6 korisnika označava da im se sviđa
    Prikaži ovu nit
    Prikaži ovu nit
    Poništi
  4. 17. pro 2019.

    <-- my most recent paper with & J. Tompson introducing *ValueDICE* - an off-policy imitation learning alg. We set new SOTA for online imitation learning, and for the 1st time (afaik) beat behavior cloning in the totally offline regime.

    1 reply 30 proslijeđenih tweetova 104 korisnika označavaju da im se sviđa
    Prikaži ovu nit
    Prikaži ovu nit
    Poništi
  5. 4. pro 2019.

    Very excited about my new paper! We formulate the on-policy max-return RL objective w.r.t *arbitrary* offline data and without *any* explicit importance correction. Amazingly, the gradient of the objective w.r.t pi is exactly the on-policy policy gradient!

    Korisnici , i
    28 proslijeđenih tweetova 121 korisnik označava da mu se sviđa
    Poništi
  6. 27. stu 2019.

    Offline RL -what do you need to know about this notoriously difficult regime? Although recent papers propose a variety of algorithmic novelties, we find many of these unnecessary in practice. Extensive studies will hopefully guide future research &practice

    16 proslijeđenih tweetova 113 korisnika označava da im se sviđa
    Poništi
  7. 8. lis 2019.

    [new paper] Adding a simple linear constraint to a loss function (commonly used in ML fairness) can have strange and unintuitive effects on the resulting model. While such constraints seem natural/harmless, we don't fully understand their consequences yet

    2 proslijeđena tweeta 15 korisnika označava da im se sviđa
    Poništi
  8. 24. ruj 2019.

    new paper! We investigate the underlying reasons for success of hierarchical RL, finding that (surprisingly) much of it is due to exploration, and that this benefit can be achieved *without* explicit hierarchies of policies

    34 proslijeđena tweeta 162 korisnika označavaju da im se sviđa
    Poništi
  9. 29. kol 2019.

    A few people had asked for this so we open-sourced our code for "Identifying and Correcting Label Bias in Machine Learning". Hope people can find it useful and build on top of it!

    2 proslijeđena tweeta 22 korisnika označavaju da im se sviđa
    Poništi
  10. 14. kol 2019.

    I'm excited to announce the release of my most recent work: applying HRL to robotic "manipulation via locomotion" tasks with impressive real-world results! w/ amazing collaborators - and more!

    7 replies 54 proslijeđena tweeta 216 korisnika označava da im se sviđa
    Poništi
  11. 29. lip 2019.

    Nice blog post!

    Near-Optimal Representation Learning for Hierarchical by Sherwin Chen cc:
    7 korisnika označava da im se sviđa
    Poništi
  12. 25. lip 2019.

    New paper out! An advancement in properly estimating off-policy occupancy ratios. We apply it to off-policy policy evaluation with great results, but we believe it should be useful in many more off-policy settings!

    7 proslijeđenih tweetova 32 korisnika označavaju da im se sviđa
    Poništi
  13. proslijedio/la je Tweet
    20. lip 2019.

    Great thanks to for the honor to kickoff the Deep Reinforcement Learning Summit with the important topic on "Secure Deep Reinforcement Learning"! Amazing talks from fellow speakers and others!

    3 proslijeđena tweeta 22 korisnika označavaju da im se sviđa
    Poništi
  14. 14. lip 2019.

    Our submission won best paper at the RL4RL workshop at ICML!

    1 reply 13 proslijeđenih tweetova 91 korisnik označava da mu se sviđa
    Poništi
  15. 9. lip 2019.

    Nice work with great collaborators - we'll be presenting this at ICML! Tue Jun 11th 03:05PM @ Room 104 and Tuesday poster #108.

    I am glad to introduce the DeepMDP! in collaboration with Saurabh Kumar, , , . We did the theory on how to learn latent space models, and it works!
    Prikaži ovu nit
    6 korisnika označava da im se sviđa
    Poništi
  16. proslijedio/la je Tweet
    24. ožu 2019.

    Fast & Simple Resource-Constrained Learning of Deep Network Structure By Suppose you have a working CNN, MorphNet will adjust number of channels in each layer to satisfy memory/latency constraints

    12 proslijeđenih tweetova 31 korisnik označava da mu se sviđa
    Poništi
  17. 22. ožu 2019.

    Open source library for "MorphNets: Fast & Simple Resource-Constrained Learning of Deep Network Structure" has been released . Based on work (CVPR 2018) with and others at ().

    1 reply 4 proslijeđena tweeta 21 korisnik označava da mu se sviđa
    Poništi
  18. 10. velj 2019.

    Great summary of our recent work about safe RL, a collaboration between researchers at Google, DeepMind, and FAIR:

    4 proslijeđena tweeta 13 korisnika označava da im se sviđa
    Poništi
  19. proslijedio/la je Tweet
    29. sij 2019.
    Odgovor korisniku/ci

    Hey, a really interesting and elegant paper! I wrote a summary of the main ideas, aiming to make them accessible to a broad audience. I hope you like it!

    1 proslijeđeni tweet 3 korisnika označavaju da im se sviđa
    Poništi
  20. 16. sij 2019.

    My new paper on learning fair machine learning classifiers: We frame the problem as trying to learn with respect to unknown (and true) labels despite only having access to observed (and biased) labels. We find a surprisingly simple solution for doing so!

    47 proslijeđenih tweetova 182 korisnika označavaju da im se sviđa
    Poništi

@ofirnachum još nije objavio nijedan Tweet.

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:

    ·