Scott Niekum

@scottniekum

Assistant professor in CS . Robotics, reinforcement learning, and imitation learning.

Vrijeme pridruživanja: veljača 2019.

Tweetovi

Blokirali ste korisnika/cu @scottniekum

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

  1. Prikvačeni tweet
    9. srp 2019.

    Want to get up to speed on what’s happening in robot learning for manipulation? Oliver Kroemer, George Konidaris, and I have just released a new survey paper: . And please let us know if there is important work we missed — it is a huge and growing field!

    Poništi
  2. proslijedio/la je Tweet
    29. sij

    Our paper 'A distributional code for value in dopamine-based reinforcement learning' on the cover of ! Read it here: Shout out to the amazing artists/designers at who make this possible, while we get to focus on the research.

    Poništi
  3. 1. velj

    Quanta discusses our recent work on risk-aware reward inference, along with great work from and others. Excited to see safe learning and value alignment problems gaining traction!

    Poništi
  4. proslijedio/la je Tweet

    : I see overwhelming discussion on the downsides of academia. I worry what this does to the hope & inspiration of grad students with ambitions to go into academia. Given that I think is the best job ever, my thoughts on life as an academic. A thread 👇🏽

    Prikaži ovu nit
    Poništi
  5. proslijedio/la je Tweet
    4. sij

    1/ Dear friends (and P.I.s), in these volatile times, do spare a thought for students of Iranian origin, especially those studying in the US. The past decade was devastating for most of us who had close ties to Iran, no matter what political background we came from.

    Prikaži ovu nit
    Poništi
  6. 18. pro 2019.

    This is great and more of us should do it. I met with a less ambitious number, but it was the highlight of NeurIPS for me. Fav part was learning about what opportunities had been most impactful in the journey of underrepresented students and the great work that they went on to do

    Poništi
  7. 18. pro 2019.

    Opinion: ML conf ACs have too much subjective tastemaking power and rejection quotas are arbitrary. Proposal: accept papers via pre-set score thresholds; the job of AC becomes convincing reviewers to move their scores, but that is the AC's only mechanism of influence. Thoughts?

    Poništi
  8. proslijedio/la je Tweet

    Simple potential solution to prevent works like these from being published in our venues: Add a question in the review form about whether the submission raises ethical concerns. If a reviewer answers yes then the paper gets a second look before a reject/review decision is made.

    Tweet je nedostupan.
    Poništi
  9. 12. pro 2019.

    Providing probabilistic guarantees for RL and imitation learning is v. challenging in data-poor settings (e.g. robotics). I'll be speaking at the NeurIPS Safety and Robustness Workshop on Friday (2:40pm) about techniques that begin to address this issue.

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

    Peer review at ML conferences.

    Poništi
  11. proslijedio/la je Tweet
    5. pro 2019.

    Really excited for next week and to present our spotlight on credit assignment :) tl;dr We can rewrite value functions in terms of a hindsight quantity that explicitly captures credit assignment and get a whole new family of RL algs! 🥳

    Prikaži ovu nit
    Poništi
  12. 4. pro 2019.

    6/ Many of these are easier said than done given the shortage of qualified reviewers / ACs. And I still think better decisions were made via this model than in many other conferences. Nonetheless, I found myself frequently frustrated during the process. Would love to hear ideas!

    Prikaži ovu nit
    Poništi
  13. 4. pro 2019.

    5/n Don't invite back reviewers that didn't respond to rebuttals: The iterative process is awesome when it works, but often some (or all!) of the reviewers don't respond, making it incredibly difficult to fairly judge the impact of the rebuttal without a huge amount of effort.

    Prikaži ovu nit
    Poništi
  14. 4. pro 2019.

    4/n Don't "calibrate" scores: ACs were sent adjusted scores based on if reviewers tended to be more positive or negative than their peers overall. IMO this just adds noise and encourages ACs to act as a threshold function rather than adjust by carefully reading reviews.

    Prikaži ovu nit
    Poništi
  15. 4. pro 2019.

    3/n Have dedicated meta-reviewers, or lower AC paper load: 22 papers is simply too many if ACs are expected to lead productive discussions and provide high quality meta-reviews, given the extensive back-and-forth ICLR model.

    Prikaži ovu nit
    Poništi
  16. 4. pro 2019.

    2/n Finer-grained scoring: Restricting scores to 1,3,6,8 led to something like half my assigned papers being 6/6/3 or 6/3/3. No one wants to give a 1 or an 8 except in extreme circumstances, so almost all granularity is lost and decision making is unnecessarily hard.

    Prikaži ovu nit
    Poništi
  17. 4. pro 2019.

    1/n There's a lot to like about the pioneering review model, but I'm feeling somewhat negative about my experience as an area chair this year. Some opinions about what could be improved:

    Prikaži ovu nit
    Poništi
  18. 3. pro 2019.

    I'm often critical of OpenAI, but this is great work that will support deeper inquiry into generalization in RL.

    Poništi
  19. proslijedio/la je Tweet
    27. stu 2019.

    If you are a US citizen or resident, (or more generally, any dept in ) offers waivers for the PhD application fee for those in need. The procedure is outlined here:

    Poništi
  20. 21. stu 2019.

    Shout out to my friend and collaborator Phil Thomas, along with and others, for their new paper in Science on a general framework for defining and avoiding undesirable behavior in ML algorithms:

    Poništi
  21. 12. stu 2019.

    Eventually Frank will be decommissioned after telling one too many off-color jokes at work. Gary Marcus will intervene to try to save him -- it wasn't his fault, it was just in his corpus. But it will be too late. We'll miss you Frank.

    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:

    ·