Michal Wolski

@michalwols

CTO @ Bite AI, teaching computers to understand food. Previously , and .

Greenpoint, NYC
Vrijeme pridruživanja: studeni 2011.

Tweetovi

Blokirali ste korisnika/cu @michalwols

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

  1. proslijedio/la je Tweet
    22. sij
    Odgovor korisnicima i sljedećem broju korisnika:

    Just to clarify some misconceptions from this thread: PyTorch has supported higher-order (reverse-mode) differentiation for years now, while forward mode and auto-vectorization (without restriction to a functional subset of Python!) is on its way!

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

    Finally can reveal our paper on ELECTRA, much more efficient than existing pretraining, state-of-the-art results; more importantly, trainable with one GPU! Key idea is to have losses on all tokens. Joint work , , .

    Prikaži ovu nit
    Poništi
  3. 18. pro 2019.

    And it looks like this: Render your html with 4gigs of js, in a docker container on kubernetes using kustomize and skaffold. !

    Prikaži ovu nit
    Poništi
  4. 18. pro 2019.

    The future is here.

    Prikaži ovu nit
    Poništi
  5. 5. pro 2019.
    Prikaži ovu nit
    Poništi
  6. 5. pro 2019.
    Prikaži ovu nit
    Poništi
  7. proslijedio/la je Tweet
    5. pro 2019.
    Odgovor korisniku/ci

    By unnoticed you mean published for just a year?

    Poništi
  8. proslijedio/la je Tweet

    Here are the results of this poll: Sanders 24% Warren 22% Biden 14% Buttigieg 12% Now look at the headline. It takes the LA Times three paragraphs to mention who is leading.

    Prikaži ovu nit
    Poništi
  9. proslijedio/la je Tweet

    That Bernie grasped this so clearly in 1991 is nothing short of remarkable. Whatever you think of , please watch this. It's hard not to contemplate all that we have lost by treating people like Sanders like a crank (and Larry Summers like a prophet). Can we stop now???

    Tweet je nedostupan.
    Poništi
  10. proslijedio/la je Tweet
    26. stu 2019.

    SuperGlue: Learning Feature Matching with Graph Neural Networks. “A neural model that simultaneously performs context aggregation, feature matching, and filtering in a single unified architecture.”

    Poništi
  11. proslijedio/la je Tweet
    26. stu 2019.

    We also introduce a technique [] for training neural networks that are sparse throughout training from a random initialization - no luck required, all initialization “tickets” are winners.

    Prikaži ovu nit
    Poništi
  12. proslijedio/la je Tweet
    22. stu 2019.

    Code release for paper on Learning Data Manipulation: Learning to augment and re-weight data in low data regime or in presence of imbalanced labels. Code: via & Bowen Tan.

    Poništi
  13. proslijedio/la je Tweet
    23. stu 2019.

    The TLDR of the paper; use adversarial examples as training data augmentation, maintain separate BatchNorm for normal vs adversarial examples. Neat. As usual I've ported & tested weights

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

    Excited to share our work on efficient neural architectures for object detection! New state-of-the-art accuracy (51 mAP on COCO for single-model single-scale), with an order-of-magnitude better efficiency! Collaborated with and .

    Poništi
  15. proslijedio/la je Tweet
    14. stu 2019.

    How Americans put up with their horrible health insurance non-system is beyond me. No one else in the developed world ever has to go through this. No one. Ever.

    Poništi
  16. proslijedio/la je Tweet
    12. stu 2019.

    Want to improve accuracy and robustness of your model? Use unlabeled data! Our new work uses self-training on unlabeled data to achieve 87.4% top-1 on ImageNet, 1% better than SOTA. Huge gains are seen on harder benchmarks (ImageNet-A, C and P). Link:

    Prikaži ovu nit
    Poništi
  17. proslijedio/la je Tweet
    7. stu 2019.

    Our new paper: Unsupervised Cross-lingual Representation Learning at Scale We release XLM-R, a Transformer MLM trained in 100 langs on 2.5 TB of text data. Double digit gains on XLU benchmarks + strong per-language performance (~XLNet on GLUE). [1/6]

    Prikaži ovu nit
    Poništi
  18. proslijedio/la je Tweet
    30. lis 2019.

    Our new pretrain model, which gives SoTA on all the generations task ( with gains of up to 6 ROUGE) while matches the performance of RoBERTa on NLU. Our 400M parameter model outperforms recent T5 770M model on NLU and 11B on CNN/DM.

    Poništi
  19. proslijedio/la je Tweet
    Odgovor korisnicima
    Poništi
  20. 24. lis 2019.
    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:

    ·