Sam Schoenholz

@sschoenholz

Research Scientist at Google Brain

Vrijeme pridruživanja: veljača 2016.
17 fotografija ili videozapisa Fotografije i videozapisi

Tweetovi

Blokirali ste korisnika/cu @sschoenholz

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

  1. Prikvačeni tweet
    6. pro 2019.

    After a ton of work by a bunch of people, we're releasing an entirely new Neural Tangents. Paper: Github: Colab Notebook:

    5 replies 130 proslijeđenih tweetova 500 korisnika označava da im se sviđa
    Prikaži ovu nit
    Prikaži ovu nit
    Poništi
  2. proslijedio/la je Tweet
    1. velj
    Odgovor korisnicima i sljedećem broju korisnika:

    cc - see this^ and also 's paper: . They get AlexNet performance with a GP!! I think this technique is widely applicable for non-parametric Bayesian inference on raw astronomical images.

    2 proslijeđena tweeta 6 korisnika označava da im se sviđa
    Poništi
  3. 28. sij

    I was particularly pleased with how easily everything came together in this colab since it more-or-less reproduces the results of () and (). Of course here it's easy to play around with the architecture and see what changes.

    1 proslijeđeni tweet 8 korisnika označava da im se sviđa
    Prikaži ovu nit
    Prikaži ovu nit
    Poništi
  4. 28. sij

    This includes a bonus-double-vmap through Newton's method that *just worked* and was super fast. Hacked on this with along with a lot of helpful comments from Roman Novak, , , and .

    1 reply 2 korisnika označavaju da im se sviđa
    Prikaži ovu nit
    Prikaži ovu nit
    Poništi
  5. 28. sij

    This is a great question that I've gotten periodically. Previously it would have taken too long to put something together, but using Neural Tangents () it's really easy and fast! Here is the reproduction in a colab:

    Can someone help me in reproducing the code for the paper - by . I'm interested in reproducing the Order Chaos Transition Graph for a Non-Linear Activation function as shown on the Figure 1 in the paper.
    18 proslijeđenih tweetova 103 korisnika označavaju da im se sviđa
    Prikaži ovu nit
    Prikaži ovu nit
    Poništi
  6. proslijedio/la je Tweet
    25. sij

    Thank you to for charging so brilliantly in what is perhaps one of the last stands Americans can do for truth, right, and democracy. I admire you.

    8:38
    Donald Trump must be convicted and removed from office. Because he will always choose his own personal interest over our national interest. Because in America, right matters. Truth matters. If not, no Constitution can protect us. If not, we are lost.
    1 reply 1 proslijeđeni tweet 24 korisnika označavaju da im se sviđa
    Poništi
  7. proslijedio/la je Tweet
    24. sij

    Research on the Neural Tangent Kernel (NTK) almost exclusively uses a non-standard neural network parameterization, where activations are divided by sqrt(width), and weights are initialized to have variance 1 rather than variance 1/width.

    23 proslijeđena tweeta 123 korisnika označavaju da im se sviđa
    Prikaži ovu nit
    Prikaži ovu nit
    Poništi
  8. proslijedio/la je Tweet
    22. sij

    Have you ever wondered what will be the ML frameworks of the '20s? In this essay, I examine the directions AI research might take and the requirements they impose, concluding with an overview of what I believe to be two strong candidates: JAX and S4TF.

    13 replies 128 proslijeđenih tweetova 521 korisnik označava da mu se sviđa
    Poništi
  9. proslijedio/la je Tweet
    20. sij

    We used JAX for competitive gradient descent (CGD) with Hessian-vector products. Mixed mode differentiation in JAX makes this efficient (just twice cost of backprop). We used CGD for training GANs and for constrained problems in RL. This library will be very useful

    Flax: A neural network library for JAX designed for flexibility (pre-release)
    1 reply 39 proslijeđenih tweetova 267 korisnika označava da im se sviđa
    Prikaži ovu nit
    Prikaži ovu nit
    Poništi
  10. proslijedio/la je Tweet
    17. sij

    Did you say more tiles? #1112

    8 replies 31 proslijeđeni tweet 229 korisnika označava da im se sviđa
    Poništi
  11. proslijedio/la je Tweet
    17. sij

    To make my research more reproducible, extensible and comparable to that of others & out of need to homogenize the language we use to express nn pruning methods, I contributed `nn.utils.prune` to 1.4 (see highlights ) Try it out & build on it! 🔥

    56 proslijeđenih tweetova 325 korisnika označava da im se sviđa
    Poništi
  12. proslijedio/la je Tweet
    15. sij

    Differentiable Digital Signal Processing (DDSP)! Fusing classic interpretable DSP with neural networks. ⌨️ Blog: 🎵 Examples: ⏯ Colab: 💻 Code: 📝 Paper: 1/

    27 replies 455 proslijeđenih tweetova 1.505 korisnika označava da im se sviđa
    Prikaži ovu nit
    Prikaži ovu nit
    Poništi
  13. proslijedio/la je Tweet
    13. sij

    pyhf 0.3.4 now supports JAX! And after JIT it's the fastest backend yet to perform particle physics hypothesis tests (even just on CPU). Thanks for some early PRs on einsum :)

    8 proslijeđenih tweetova 27 korisnika označava da im se sviđa
    Poništi
  14. proslijedio/la je Tweet
    12. sij

    The Case for Bayesian Deep Learning ”Bayesian or not, the prior will certainly be imperfect. Avoiding an important part of the modelling process because one has to make assumptions, however, will often be a worse alternative than an imperfect assumption.”

    127 proslijeđenih tweetova 542 korisnika označavaju da im se sviđa
    Poništi
  15. proslijedio/la je Tweet
    10. sij

    Reality and simulation converging...

    0:29
    I have added some dynamic lighting to my VR Train scene. All made inside and rendered with
    Prikaži ovu nit
    5 replies 32 proslijeđena tweeta 232 korisnika označavaju da im se sviđa
    Poništi
  16. proslijedio/la je Tweet
    10. sij

    1/2 Why isn't it more common to do explicit Hamiltonian MCMC on a Bayesian Neural Network's weights, with eg the initial condition = the loss minima found via SGD? I'm playing around with one in JAX and it seems to be working reasonably even with 5 chains:

    9 replies 22 proslijeđena tweeta 171 korisnik označava da mu se sviđa
    Prikaži ovu nit
    Prikaži ovu nit
    Poništi
  17. proslijedio/la je Tweet
    9. sij

    What I did over my winter break! It gives me great pleasure to share this summary of some of our work in 2019, on behalf of all my colleagues at & .

    11 replies 328 proslijeđenih tweetova 1.168 korisnika označava da im se sviđa
    Prikaži ovu nit
    Prikaži ovu nit
    Poništi
  18. proslijedio/la je Tweet
    8. sij

    Fenchel-Rockafellar duality is a powerful tool that more people should be aware of, especially for RL! Straightforward applications of it enable offpolicy evaluation, offpolicy policy gradient/imitation learning, among others

    31 proslijeđeni tweet 108 korisnika označava da im se sviđa
    Poništi
  19. proslijedio/la je Tweet
    6. pro 2019.

    Our review of Machine Learning and the Physical Sciences made the cover of Review of Modern Physics ! (or here ) I worked on that image for a while and it incorporates the famous CNN figure from

    Korisnici , , i
    41 proslijeđeni tweet 162 korisnika označavaju da im se sviđa
    Prikaži ovu nit
    Prikaži ovu nit
    Poništi
  20. proslijedio/la je Tweet
    2. sij
    Odgovor korisnicima

    e & q = tuning forks γ= phonons g= ping pong ball

    1 proslijeđeni tweet 55 korisnika označava da im se sviđa
    Poništi
  21. proslijedio/la je Tweet
    14. pro 2019.

    We used mixed mode differentiation in JAX to implement competitive gradient descent that requires Hessian-vector products. Code repo:

    1 reply 8 proslijeđenih tweetova 25 korisnika označava da im se sviđa
    Prikaži ovu nit
    Prikaži ovu nit
    Poništi

@sschoenholz još nije objavio nijedan Tweet.

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:

    ·