Brandon Rohrer

@_brohrer_

Machine learning. Data science. Robotics. he/him. Opinions are just mine. iRobot. ex-Facebook, Microsoft, DuPont, Sandia. MIT.

Boston
Vrijeme pridruživanja: kolovoz 2012.

Tweetovi

Blokirali ste korisnika/cu @_brohrer_

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

  1. Prikvačeni tweet
    16. ruj 2019.

    Backpropagation is the beating heart of neural networks. Here's how it works.

    How Backpropagation Works. with diagram of pipes and update equations.
    Prikaži ovu nit
    Poništi
  2. prije 3 sata

    Excellent context and a fascinating read from .

    Poništi
  3. prije 7 sati

    I'm teaching an Intro to Deep Learning workshop sequence at East in April! Come join me for a full day of concepts and code, no prior experience required.

    Poništi
  4. proslijedio/la je Tweet
    prije 14 sati

    The iRobot Comms team is hiring. Do you know someone in the Boston area with 1-2 years of comms experience who is excited about consumer tech, robotics, media relations, reviews and writing? I am proud to say that we have a great team.

    Poništi
  5. prije 14 sati

    Details of my implementation and some things I learned:

    Prikaži ovu nit
    Poništi
  6. prije 17 sati

    StatQuest by is one of the best statistics and ML resources in existence. He's now committed to making tutorials for us full-time. StatQuest shows what educational internet can be in the right hands.

    Poništi
  7. prije 18 sati

    Actual moral of the story: keep building, keep asking why and how. 11/11

    Prikaži ovu nit
    Poništi
  8. prije 18 sati

    Most importantly, I understand so much more now than I would have if I had skipped right to the authors’ solution. Re-implementation gives you understanding. Re-derivation gives you a sense of ownership. 10/11

    Prikaži ovu nit
    Poništi
  9. prije 18 sati

    Just kidding. My data was different than the original authors’ so my results were pretty different too. I did some small, but I think interesting, things differently. 9/11

    Prikaži ovu nit
    Poništi
  10. prije 18 sati

    Moral of the story: everything interesting has already been done 8/11

    Prikaži ovu nit
    Poništi
  11. prije 18 sati

    I was thrilled by the result. I went to share it and a more careful search showed that the method is already been presented very thoroughly in a 2013 arxiv paper: k-Sparse Autoencoders by Alireza Makhzani and Brendan Frey 7/11

    Screenshot of paper in link: k-sparse autoencoders
    Prikaži ovu nit
    Poništi
  12. prije 18 sati
    Prikaži ovu nit
    Poništi
  13. prije 18 sati

    With sparsification, these nodes have much more structure and form a reasonable sparse basis. 5/11

    Representation of hidden nodes in a sparse layer, most cases showing structure. Includes horizontal, vertical, diagonal lines of varying spatial frequencies, blobs, and checkerboards.
    Prikaži ovu nit
    Poništi
  14. prije 18 sati

    Without sparsification, these nodes are information rich, but visually uninteresting. They look like static. 4/11

    Representation of hidden nodes in a dense network. Each node shows a wide variety of values with no clear pattern.
    Prikaži ovu nit
    Poništi
  15. prije 18 sati

    Using it in an autoencoder lets you see what it does. You can visualize nodes in the bottleneck layer by finding what output image they produce if all the other nodes are zero. 3/11

    Prikaži ovu nit
    Poništi
  16. prije 18 sati

    So I made a layer that preserved the k node activities with the highest magnitudes. The rest it set to zero. 2/11

    Prikaži ovu nit
    Poništi
  17. prije 18 sati

    A research story with a twist I looked for a neural network regularization method that limited the number of non-zero node activities (L0 group sparsity?) I couldn’t find one. 1/11

    Prikaži ovu nit
    Poništi
  18. 3. velj
    Line drawing of a four-armed robot washing the dishes
    Prikaži ovu nit
    Poništi
  19. 3. velj

    I figured out how to solve the problem of cheesy robot stock photos: Cheesy robot drawings

    Prikaži ovu nit
    Poništi
  20. 3. velj

    Also (new ideas) (decorate your laptop) As always you are invited to fork my library repository () and use it as a template for your own blog.

    Prikaži ovu nit
    Poništi
  21. 3. velj

    I've included some handy shortcuts too (course listing) (tutorial posts) (Office Hours blog) (Course 312. Find courses directly by number)

    Prikaži ovu nit
    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:

    ·