Pierre-Yves Oudeyer

@pyoudeyer

Research director , Head of lab. Studying autonomous learning and curiosity in AI, machine learning, cognitive sciences, and education.

Vrijeme pridruživanja: rujan 2013.

Tweetovi

Blokirali ste korisnika/cu @pyoudeyer

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

  1. Prikvačeni tweet
    19. stu 2018.
    Prikaži ovu nit
    Poništi
  2. prije 4 sata

    Combining it with a setup like the one of 's paper would be very interesting, both in terms of fundamental research and in terms of making new discoveries in chemistry.

    Prikaži ovu nit
    Poništi
  3. prije 4 sata

    We investigated this in our 2020 paper with the IMGEP-OGL algorithm, to discover novel "animal" forms in Lenia, a continuous game of life from Intrinsically Motivated Discovery of Diverse Patterns in Self-Organizing Systems

    Prikaži ovu nit
    Poništi
  4. prije 4 sata

    One scientific challenge in this area is how to learn useful representations of goal spaces while exploring and making discoveries in such high-dimensional systems (e.g. in these wet experiments, the goal space was engineered by the hand, but how to learn it from pixels?)

    Prikaži ovu nit
    Poništi
  5. prije 4 sata

    This extends the Explauto library: dedicated to population-based IMGEP algorithms, developed by and

    Prikaži ovu nit
    Poništi
  6. prije 4 sata

    The code used in et al.'s paper is available here:

    Prikaži ovu nit
    Poništi
  7. prije 4 sata

    The particular IMGEP algorithm used in these wet experiments samples goals randomly, leveraging a learned forward model + stochastic optimization to infer the corresponding params. to try. It is described in (wth L-BFGS instead of CMA-ES)

    Prikaži ovu nit
    Poništi
  8. prije 4 sata

    Great work showing how intrinsically motivated goal exploration can be used to automate discovery of novel patterns in wet chemical systems !!! Congratulations and team! Intrinsically motivated algorithms open new perspectives in physics/chemistry

    Prikaži ovu nit
    Poništi
  9. proslijedio/la je Tweet
    30. sij

    1/ Did you know that Vannevar Bush (you know, the guy who helped enable everything from radar to the manhattan project, the NSF to memexes) wrote an autobiography? Turns out that yes he did, it's been out of print since the 70's, and it's *excellent* BOOK REPORT THREAD

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

    There is also this great blog post by the same author (Cedric Colas) that's a bit more accessible:

    Poništi
  11. 30. sij

    Amazing: every week I see a paper comparing algorithms using mean performance over *3* seeds ! Yes, ****3**** !!! Please please community, your great ideas will be served better using standard scientific methods!

    Poništi
  12. proslijedio/la je Tweet
    28. sij

    This is such a useful resource for understanding AI Ethics and Safety and it shows so clearly that AI is a multi-disciplinary field.

    Poništi
  13. proslijedio/la je Tweet
    29. sij

    RL! What is it good for? Actually... Real world robotics!

    Poništi
  14. 23. sij

    Thanks to all contributors and to for helping me editing the newsletter!

    Prikaži ovu nit
    Poništi
  15. 23. sij

    I invite all readers interested to participate to this dialog to send me their response by June 30th, 2020. The length of each response must be between 600 and 800 words including references.

    Prikaži ovu nit
    Poništi
  16. 23. sij

    However, a key challenge is how to measure progress in this area, as traditional RL benchmarks were not constructed to address open-ended learning.

    Prikaži ovu nit
    Poništi
  17. 23. sij

    Another related line of work is open-ended evolution in multi-agent systems with co-evolution.

    Prikaži ovu nit
    Poništi
  18. 23. sij

    Several steps in this direction have been made in the last decade, such as algorithms enabling agents to imagine their own goals and self-organize their learning curriculum (initially in the CDS community, and recently extended in the machine learning community).

    Prikaži ovu nit
    Poništi
  19. 23. sij

    While impressive progress was made recently in reinforcement learning, as shown in benchmarks with a set of pre-defined external objectives, a major challenge is now to build autonomous agents that can progressively discover and learn open repertoires of skills in open worlds.

    Prikaži ovu nit
    Poništi
  20. 23. sij

    Then, a new dialog initiation is proposed by on a fundamental question for the future of AI: How to evaluate open-ended learning agents?

    Prikaži ovu nit
    Poništi
  21. 23. sij

    They also offer a perspective on how artificial intelligence can play a key role in impactful educational apps, enabling to implement personalized and motivating learning strategies

    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:

    ·