Michaël Defferrard

@m_deff

researcher interested in and . PhD student of . Teaching Data Science at . Open science, source, data.

EPFL, Switzerland
Vrijeme pridruživanja: svibanj 2015.

Tweetovi

Blokirali ste korisnika/cu @m_deff

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

  1. Prikvačeni tweet
    12. tra 2019.

    While they can't image black holes (yet?), graph NNs can help us research the cosmos! DeepSphere: a graph-based spherical CNN with approximate equivariance. Joint work with and R. Sgier from . Story in 6 tweets 👇

    Prikaži ovu nit
    Poništi
  2. prije 13 sati

    Spinach is not a good source of iron. Amazing how myths persist when we don't pay attention (it was debunked in the 1930s). If you believed the myth from Popeye (as I did), do also know that its author chose spinach for vitamin A, not iron.

    Poništi
  3. 29. sij

    I have a secret smile whenever I pause to admire the marvel of engineering that makes my computer work, from web apps to transistors.

    Poništi
  4. 29. sij

    Is spinach a rich source of iron?

    Poništi
  5. 28. sij

    That's what focus brings. Free of bureaucracy, committees, teaching, supervising, grant writing, etc., researchers can research.

    Poništi
  6. 13. sij

    In Europe (at least in 🇨🇭 and around), trains are more comfortable and not much slower than planes and cars. Both for local and long-distance trips (e.g., I took a 6h direct train from Basel to Amsterdam). I'm concerned about depending more on planes and cars if moving to SF. :/

    Poništi
  7. 5. sij

    Just finished Bad Blood by . It's infuriating how far you can get with plain lies. :/ Hopefully the majority will too see the truth in Trump and Brexiters, and they'll face a similar ending. PS: the book reads like a thriller. Highly recommended!

    Poništi
  8. 1. sij

    We spend so much time discussing issues --- which is necessary to progress --- that we forget where we come from. Feels good to take some time to appreciate it. :) And is fantastic!

    Poništi
  9. 29. pro 2019.

    The universe is dangerous. Yet we don't perceive it and discover slowly. Most likely solution (the rare environment) to Fermi's paradox?

    Poništi
  10. proslijedio/la je Tweet
    27. pro 2019.

    New work "what graph neural networks cannot learn: depth vs width" (ICLR20) studies the expressive power of GNN: It provides sufficient conditions for universality and shows that many classical problems are impossible when depth x width < c n^p. Blogpost:

    Prikaži ovu nit
    Poništi
  11. 23. pro 2019.

    There's a huge business opportunity in facilitating the extraction of relevant information from data.

    Poništi
  12. 21. pro 2019.

    While Lisp's axiomatization of computing is intellectually beautiful and powerful, Lisp code is hard to read. I believe a more literate Lisp would go a long way towards adoption. And that would be a huge step forward for computing!

    Poništi
  13. 20. pro 2019.

    Not surprised that UMAP's preservation of global structure is entirely due to initialization. Force layouts (used by UMAP & t-SNE) do local updates only. A contrario, Laplacian eigenmaps and PCA use the global structure. Init globally and fine-tune locally is the wining strategy.

    Poništi
  14. 12. pro 2019.

    What is so special about gender that it's ingrained in our languages? English would be as comprehensible with only "they" as a pronoun as with "she", "he", "they" (same with "their" and other languages). Why did it evolved gendered?

    Poništi
  15. 11. pro 2019.

    Checkout 's tutorial if you ever wondered how to make beautiful graphs and publish an interactive visualization on the web!

    Poništi
  16. 9. pro 2019.

    This year we're trying peer-review for students to internalize grading criteria and give constructive feedback. Got promising proposals!

    Poništi
  17. 26. stu 2019.

    BTW, you can view the first term A^0 = I as skip connections.

    Prikaži ovu nit
    Poništi
  18. 26. stu 2019.

    GCN can only do low-pass filters because of the self-loops in the adjacency matrix that mix the 0- and 1-neighborhood. Guess what is the simplest fix? Use a polynomial. Even a first-order Y = X W_0 + A X W_1 (and that is ChebNet with K=2).

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

    Nice article comparing features across some cloud services for reproducibility. Wish they spent more time talking about the means of production for these projects, which I see as a key differentiator and a big part of sustainability for open scholarship

    Prikaži ovu nit
    Poništi
  20. 29. lis 2019.
    Poništi
  21. 21. lis 2019.

    Switzerland 🇨🇭 elected its parliament for the next 4 years. Good in my opinion: more women (+10%), more green parties (+9%, mostly at the expense of the right), good balance. Bad: participation of 45% only (-3.5%). More at &

    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:

    ·