Matt Henderson

@matthen2

maths, conversational AI. lead scientist , previously . living in Singapore

Vrijeme pridruživanja: veljača 2010.

Tweetovi

Blokirali ste korisnika/cu @matthen2

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @matthen2

  1. 31. sij

    Work with me this summer on conversational AI at PolyAI Singapore

    Poništi
  2. proslijedio/la je Tweet
    13. sij

    Excited to share our work NADST: The first Non-Autoregressive neural model for Dialogue State Tracking (DST) w/ SOTA on MultiWOZ2.1, reducing inference latency by an order (). Joint work w/ and at !

    Poništi
  3. 8. sij

    got around to trying it with Tacotron

    Prikaži ovu nit
    Poništi
  4. 14. pro 2019.
    Poništi
  5. proslijedio/la je Tweet
    29. stu 2019.

    Have you tried the new ConveRT featurizer yet? We're keen to see how the new pipeline performs on your dataset. Join the discussion in the community forum and contribute your evaluation numbers

    Poništi
  6. 26. stu 2019.

    Our ConveRT model has been implemented in Rasa, making it very easy to use as a feature extractor for intent classification

    Poništi
  7. 15. stu 2019.

    it is working a lot better than other attention mechanisms I've tried for my task-

    Prikaži ovu nit
    Poništi
  8. 14. stu 2019.

    their prior filter, convolved on the previous alignment, helps guide the model to alignments that move forward in time code for animation

    Prikaži ovu nit
    Poništi
  9. 14. stu 2019.

    this is a great paper with an elegant idea for learning monotonic attention that generalises to long sequences

    Prikaži ovu nit
    Poništi
  10. 12. stu 2019.

    ConveRT results on DSTC7 can be reproduced using the code in the repo:

    Poništi
  11. 11. stu 2019.
    Prikaži ovu nit
    Poništi
  12. 11. stu 2019.

    The efficient ConveRT (Conversational representations from transformers) model is available as a hub module at:

    Prikaži ovu nit
    Poništi
  13. 11. stu 2019.

    . has released ConveRT, a pretrained sentence encoder optimized for conversational understanding. It's under 60MB, is substantially faster & cheaper to train than standard encoder models (BERT etc.), and beats those models on conversational tasks:

    Prikaži ovu nit
    Poništi
  14. proslijedio/la je Tweet
    5. stu 2019.

    This never ends. This year, so far, 15 out of 44 people to attend workshop at (which is still in Canada) have been denied visas. That's 33%. We had all this press last year, they were supposed to help us this year.

    Prikaži ovu nit
    Poništi
  15. proslijedio/la je Tweet
    29. lis 2019.

    Here's to hoping 2020 will bring 10x as efficient models instead of 10% reductions in GLUE error rates :)

    Poništi
  16. proslijedio/la je Tweet
    28. lis 2019.

    Super excited to be organizing the 2nd NLP For Conversational AI Workshop held at . If you like Seattle coffee ☕, Washington hikes ⛰️, and talking machines 🤖, we look forward to seeing you there!

    Prikaži ovu nit
    Poništi
  17. 28. lis 2019.

    drawing hyperbolic fractals with the chaos game in hyperbolic space

    Poništi
  18. 27. lis 2019.
    Prikaži ovu nit
    Poništi
  19. 27. lis 2019.

    how could you draw a sierpinski triangle in the hyperbolic disk? one way is to start with some points, and repeatedly move each one half way to a random one of the corners, using the hyperbolic geometry. when the corners are close enough, you get the original sierpinski

    Prikaži ovu nit
    Poništi
  20. proslijedio/la je Tweet
    26. lis 2019.

    Thanks to , I got to test out my new candidate DeOldify model on a use case that I think will be quite appealing for many people: wedding photos! Thanks Matt!

    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:

    ·