Jacob Buckman

@jacobmbuckman

PhD student at studying deep reinforcement learning. Formerly at , a resident, and .

Vrijeme pridruživanja: prosinac 2016.

Tweetovi

Blokirali ste korisnika/cu @jacobmbuckman

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

  1. 25. sij
    Poništi
  2. proslijedio/la je Tweet
    22. sij

    The whole thread on BNNs and blog post by and reminded me of the "First, you rob a bank..." characterization by Yasser Abu Mostafa Apologies to my Bayesian friends who may find it unfair.

    Poništi
  3. proslijedio/la je Tweet
    13. sij

    New blog post! This time I'm looking at recent advances in memory-efficient training and where that might lead.

    Poništi
  4. 22. sij

    The old version will remain accessible online for posterity, but I've removed it from my blog's index, and added a link to the updated version of the post.

    Prikaži ovu nit
    Poništi
  5. 22. sij

    We've updated this blog post in response to community discussion & feedback:

    Prikaži ovu nit
    Poništi
  6. 22. sij

    Updated blog post with -- "Bayesian Neural Networks Need Not Concentrate": Thanks to all for the discussion & feedback. This improved version of the blog post hopefully explains the core claims more clearly, while being less polarizing.

    Prikaži ovu nit
    Poništi
  7. 20. sij

    If a result relevant to our discussion is included in some prior work, by all means reference it! But blanket dismissal of all ideas that come from people who haven't already completed your exact personal "required reading" is a disturbingly elitist attitude.

    Prikaži ovu nit
    Poništi
  8. 20. sij

    It's frustrating when people refuse to have idea-level discussions on the grounds that "since you missed this 1 reference, you aren't worth talking to." Feels very patronizing and anti-good-discourse.

    Prikaži ovu nit
    Poništi
  9. 20. sij

    We'll update the blog post to include this new evidence. If anyone is interested in empirically investigating these questions about BNN priors in more detail, feel free to reach out and/or pull my repo!

    Prikaži ovu nit
    Poništi
  10. 20. sij

    Overall, there does seem to be a significant difference between the prior probability of just clean data, and clean + corrupted! This is evidence that BNN priors are *not* entirely generalization-agnostic. My hypothesis was wrong!

    Prikaži ovu nit
    Poništi
  11. 20. sij

    Results are in! Averaged over 3 trials: unregularized: logp(clean) = -1654049 logp(clean+corrupted) = -1654101 with .01 reg towards prior: logp(clean) = -1654024 logp(clean+corrupted) = -1654066 code:

    Prikaži ovu nit
    Poništi
  12. 18. sij

    This is a more nuanced elaboration on Carles's controversial thread from a few weeks ago:

    Prikaži ovu nit
    Poništi
  13. 18. sij

    New blog post with -- "A Sober Look at Bayesian Neural Networks": Without a good prior, Bayesian uncertainties are meaningless. We argue that BNN priors are likely quite poor, and concretely characterize one specific failure mode.

    Prikaži ovu nit
    Poništi
  14. proslijedio/la je Tweet

    I propose a grand challenge for dextrous robotic manipulation: cook eggs in all the ways shown in this video

    Prikaži ovu nit
    Poništi
  15. 2. sij
    Prikaži ovu nit
    Poništi
  16. 2. sij

    Is causal inference the same thing as "batch RL" (learning from data without interacting)? Or is there more to it?

    Prikaži ovu nit
    Poništi
  17. 22. pro 2019.

    in case anyone was wondering what it's like to argue with irl

    Poništi
  18. 22. pro 2019.

    Highly persuasive rant. Any Bayesians out there willing to defend the reverend's honor?

    Poništi
  19. proslijedio/la je Tweet

    First airplane with no moving parts: no turbines, propellers, nor jets. Uses electroaerodynamic propulsion, where 40k volts generate ions moving from a larger electrode to a smaller one, colliding with air molecules on the way to produce "ionic wind".

    Poništi
  20. proslijedio/la je Tweet
    6. pro 2019.

    Peer review at ML conferences.

    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:

    ·