Warith HARCHAOUI

@warith_h

French Data Scientist and Ph.D. student in Machine Learning

Paris, France
harchaoui.org/warith
Joined July 2014

You may also like

·

Tweets

You blocked

Are you sure you want to view these Tweets? Viewing Tweets won't unblock

  1. Retweeted
    Aug 27

    Hyper-parameter autotuning for fastText: get a 1MB text classifier while having a coffee. With this new feature, it is possible to constrain the size of the final model, and automatically find the hyper-parameters giving the best results on a validation set.

    0:14
    Facebook AI researchers are releasing a new feature for the fastText library which provides hyper-parameter autotuning for more efficient text classifiers.
    3 retweets 22 likes
  2. May 3

    Well said!

    Here is a real use case from work for model improvement and the steps taken to get there: - Baseline: 53%…
  3. Retweeted
    Apr 29

    How do you compute with an infinitely wide deep net (eg, AlexNet or VGG with width taken to infinity)? Despite crazy overparametrization, this net works OK on finite dataset CIFAR10. To understand how this was done ( via "Neural Tangent Kernels") see

    4 replies 88 retweets 406 likes
  4. Retweeted
    Apr 29

    Conférence de presse / lancement du Côte d’Azur !

    3 retweets 13 likes
  5. Apr 5

    Small post about small data:

  6. Feb 26

    Surpris que mon introduction aux réseaux de neurones ait un peu de succès 🙂

    1 like
  7. Retweeted
    Feb 22

    This morning : going to work at and seeing this !

    1 retweet 9 likes
  8. Retweeted
    1 Nov 2018

    Just a couple of years ago, I found it hard to believe that vision people were still all working with rectangular bounding boxes. I guess they’ve fixed that now. 🙂

    Today we are releasing Mask R-CNN Benchmark: a fast and modular implementation for Faster R-CNN and Mask R-CNN written entirely in 1.0. It brings up to 30% speedup compared to mmdetection during traini…
    3 replies 53 retweets 264 likes
  9. Retweeted
    11 Sep 2018

    Our paper with on the statistical properties of deep generative models “Leveraging the Exact Likelihood of Deep Latent-Variable Models” () was accepted at 😃 To celebrate, here’s a short thread with our main findings (1/6) ⬇️

    1 reply 12 retweets 51 likes
    Show this thread
    Show this thread
  10. Retweeted
    7 Sep 2018

    Mon livre Introduction au à destination d'élèves de M1/2A d'école d'ingé va bientôt sortir ! Il me reste quelques exemplaires de promotion, manifestez-vous !

    11 replies 5 retweets 36 likes
  11. Retweeted
    16 Aug 2018

    Countdowns to top CV/NLP/ML/Robotics/AI conference deadlines

    1 reply 33 retweets 82 likes
  12. Retweeted
    14 Aug 2018

    Some works of the Gauss prize winners.

    40 retweets 132 likes
  13. Retweeted
    12 Aug 2018

    David Donoho just god the Gauss prize at ICM 2018! IMHO its four main contributions: soft thresholding, basis pursuit, curvelets and compressed sensing.

    3 replies 46 retweets 200 likes
  14. Retweeted
    8 Aug 2018

    A Riemannian manifold is locally an Euclidean space. An embedded surface is an example. Defines a Riemannian tensors field (first fundamental form) in parametric space which encodes length deformations.

    71 retweets 226 likes
  15. Retweeted
    9 Aug 2018

    🎙 (Ré)écouter : "A quand un Nobel pour une intelligence artificielle ?" via ▶ avec Laurence Devillers de

    4 retweets 8 likes
  16. Retweeted
    20 Jul 2018

    Unsupervised pre-training is starting to work quite well for learning visual representations. Here is a new paper from FAIR-Paris that demonstrates how this works.

    3 replies 189 retweets 548 likes
  17. Retweeted
    18 Jul 2018

    The answer :)

    4 retweets 10 likes
    Show this thread
    Show this thread
  18. Retweeted
    9 Jul 2018

    Gan-level samples from a likelihood model + excellent unsupervised representations

    0:04
    Synthesizing realistic high-resolution images with Glow, a new reversible generative model:
    50 retweets 186 likes
  19. Retweeted
    7 Jul 2018

    Looking forward to some great tutorial presentations at next Tuesday in Stockholm! The full list is here:

    14 retweets 46 likes
  20. Retweeted
    7 Jul 2018

    Clearest derivation of duality I've seen so far (from 's )

    1 reply 5 retweets 50 likes
  21. Retweeted
    6 Jul 2018

    Très heureux de rejoindre l'équipe de statistique du @LaboMAP5 de à la rentrée de septembre !

    2 retweets 18 likes
  22. Retweeted
    5 Jul 2018

    The European commission contracted Elsevier to monitor open science in Europe. This is like putting the mafia in charge of monitoring the police and the justice system. Elsevier has historically done all it can to...

    20 replies 442 retweets 854 likes
  23. Retweeted
    5 Jul 2018

    Leon Bottou.... What doesn’t work is important to understand.

    8 retweets 22 likes
  24. Retweeted
    30 Jun 2018

    most common neural net mistakes: 1) you didn't try to overfit a single batch first. 2) you forgot to toggle train/eval mode for the net. 3) you forgot to .zero_grad() (in pytorch) before .backward(). 4) you passed softmaxed outputs to a loss that expects raw logits. ; others? :)

    102 replies 1,105 retweets 4,404 likes
    Show this thread
    Show this thread
  25. Retweeted
    28 Jun 2018

    we're getting a sneak preview of talk for on visual search . She's speaking on Monday

    4 retweets 5 likes
  26. Retweeted
    12 Jun 2018

    Neural Kernel Networks: a differentiable framework for compositional kernel learning. Like the Automatic Statistician, but trainable with gradient-based optimization rather than discrete search. By et al.

    3 replies 90 retweets 304 likes
  27. Retweeted
    10 Jun 2018

    Amusant et bien vu.

    [NoComment]
    2 retweets 9 likes
  28. Retweeted
    8 Jun 2018

    Excited to announce the FashionGen Text to Image challenge. The task is to generate fashion items based on design descriptions. Data is annotated by designers, with multiple angles of each item and has over 300,000 samples!

    , and
    10 replies 123 retweets 300 likes
  29. Retweeted
    31 May 2018

    Important message from Isabel Schwende at : the best state-of-the-art models aren't necessarily best in the wild, where you need something fast and easy to deploy.

    2 retweets 3 likes
  30. Retweeted
    31 May 2018

    Very interesting essay by Aaron Hertzmann about the role of technology in art and on the question of whether computers coukd be artists. Aaron argues that art is a social activity, and therefore machine can't really...

    1 reply 26 retweets 79 likes
  31. Retweeted
    30 May 2018

    A research team led by Hugh Herr has developed the AMI, a new surgical procedure and neural interface that gives amputees precise, natural control of their prosthetic limbs.

    5 replies 129 retweets 241 likes
    Show this thread
    Show this thread
  32. Retweeted
    30 May 2018

    Seriously?? "[A person] told Assistant to "activate gun." But in a world where machine learning has led to AI that's smart enough to anticipate our needs, it's not hard to imagine digital assistants kill off people who upset us, if given access to weapons"

    3 replies 6 retweets 16 likes
    Show this thread
    Show this thread
  33. Retweeted
    30 May 2018

    New paper on unsupervised mapping of word vectors, using Procrustes in Wasserstein distance, available on arxiv: . With and Q. Berthet. More resources to come soon!

    2 replies 31 retweets 70 likes
  34. Retweeted
    29 May 2018

    Hourra ! pour le héros

    32 replies 1,213 retweets 2,185 likes
  35. Retweeted
    29 May 2018

    Double blind peer review: it works.

    Our article on the effectiveness of anonymization in double-blind review is out! TL;DR - it’s really effective, undermining a key argument against double-blind review. (Co-authored by , Yuriy Brun, , , and Safraz Khurshid) …
    Show this thread
    2 retweets 7 likes
  36. Retweeted
    24 May 2018

    Arriving in Sicilia for the on clustering and classification!

    1 retweet 9 likes
  37. Retweeted
    23 May 2018

    Habilitation defense of Gael Gael is defending his habilita...

    4 retweets 26 likes
  38. 23 May 2018

    Congrats

    Habilitation defense of Gael Gael is defending his habilita...
  39. Retweeted
    22 May 2018
    Replying to

    The wake-sleep algorithm for unsupervised neural networks:

    1 retweet 3 likes