Maha Elbayad

@melbayad

PhD student at Inria/LIG (Grenoble university, France).

Vrijeme pridruživanja: rujan 2012.

Tweetovi

Blokirali ste korisnika/cu @melbayad

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

  1. Prikvačeni tweet
    20. pro 2019.

    Happy to share that my internship work "Depth-adaptive Transformer" has been accepted to . TL;DR: We dynamically adjust the computation per input and match the accuracy of a baseline Transformer with only 1/4 the decoder layers.

    Prikaži ovu nit
    Poništi
  2. 20. pro 2019.

    Joint work with Jiatao ( ), Edouard () & my host Michael Auli.

    Prikaži ovu nit
    Poništi
  3. proslijedio/la je Tweet
    12. pro 2019.

    Our work on FlauBERT and FLUE (language models and evaluation benchmark for French) have been released today (198th birthday of Gustave Flaubert). Paper: Code and models:

    Prikaži ovu nit
    Poništi
  4. proslijedio/la je Tweet
    12. pro 2019.

    Here is FlauBERT: a French LM learnt (with J-Zay supercomputer) on a large and heterogeneous corpus. Along with it comes FLUE (evaluation setup for French NLP). FlauBERT was successfully applied to complex tasks (NLI, WSD, Parsing). More on

    Poništi
  5. proslijedio/la je Tweet
    26. srp 2019.

    Application to 2019 PAISS summer school is now open ! October 3>5 in Paris co-organized by MIAI and PRAIRIE @u_grenoblealpes

    Poništi
  6. proslijedio/la je Tweet
    1. srp 2019.

    Available position on incremental seq2seq mapping for speech generation. Funded by the new Research Institute ( MIAI). More info at

    Poništi
  7. proslijedio/la je Tweet
    16. lip 2019.

    Interested to work on incremental sequence-to-sequence mapping for speech generation using deep neural networks (PhD funded by Grenoble Artificial Intelligence Institute) ? Please contact ( ; and laurent.girin@gipsa-lab.fr) to know more...

    Poništi
  8. proslijedio/la je Tweet
    15. lis 2018.

    I've spent most of 2018 training models that could barely fit 1-4 samples/GPU. But SGD usually needs more than few samples/batch for decent results. I wrote a post gathering practical tips I use, from simple tricks to multi-GPU code & distributed setups:

    Prikaži ovu nit
    Poništi
  9. proslijedio/la je Tweet
    26. kol 2018.

    New seq2seq architecture - jointly encodes source and targets into a 2D ConvNet. No enc/dec or explicit attention. Outperforming ConvS2S and Transformers on IWSLT'14 de<->en, with 3 to 8 times less parameters from and team

    Poništi
  10. 28. srp 2018.

    Our paper "2D Convolutional Input-Output Coding for Sequence-to-Sequence Prediction" has been accepted . Pre-print soon

    Poništi
  11. proslijedio/la je Tweet
    23. ožu 2018.

    Simplicity is still competitive! Standard LSTM/QRNN models can achieve state-of-the-art results on character (PTB, enwik8) + word level (WikiText-103) language modeling datasets in 12-48 hours w/ a single GPU. Joint work w/ &

    Character-level Penn Treebank results
    Character-level enwik8 results
    WikiText-103 results
    Analysis of hyper-parameter sensitivity
    Prikaži ovu nit
    Poništi
  12. proslijedio/la je Tweet
    28. ožu 2018.

    One week left to apply to , our summer school on AI, co-organized by , ! - July 2-6th in Grenoble. Featuring , A. Zisserman, L. Bottou, L. Agapito, M. Hebert, E. Dupoux and more.

    , , i još njih 3
    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:

    ·