Rich Harang

@rharang

ML, security, stats, awful puns, rants about bias in ML. Research Director . Also and . He/him. Personal account, views my own.

Lost in Hilbert space.
Vrijeme pridruživanja: rujan 2010.

Tweetovi

Blokirali ste korisnika/cu @rharang

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

  1. Prikvačeni tweet
    24. sij

    Debugging deep learning models can be really tricky and frustrating, especially in the security space where a lot of the time you're not sure about ground truth labels, but here's a thread with some tricks I've picked up. Add your own if you've got them! 1/

    Prikaži ovu nit
    Poništi
  2. proslijedio/la je Tweet
    5. velj

    If you wondered why AV has false positives, take a look at this excel file: 1. Runs code automatically on opening 2. Uses native APIs that manipulate memory 3. Instantiates a COM object defined in a text string in a macro Benign.

    Poništi
  3. proslijedio/la je Tweet
    prije 16 sati
    Odgovor korisniku/ci
    Poništi
  4. proslijedio/la je Tweet
    4. velj

    Our latest paper, ☢️ Radioactive data: tracing through training, is now on arXiv. TL;DR: you can modify your data in an imperceptible way so that any model trained on it will have an identifiable mark. (1/7)

    Prikaži ovu nit
    Poništi
  5. 5. velj

    Me: What absolute monster of a human being set their tab stop to *3* in this code? Also me: ...oh.

    Poništi
  6. proslijedio/la je Tweet
    4. velj
    Poništi
  7. proslijedio/la je Tweet
    4. velj

    Monster-size models are *one* possible way to go about getting better results. Green AI proposes to pay attention not only to the optimization metric but the cost of getting at it. Smaller, faster, more data-efficient models are just better.

    Poništi
  8. 4. velj

    Today's "I need coffee moment": getting really mad that I couldn't import a module that was obviously in my conda environment, before realizing that it was *ipython* that wasn't in my conda env and I was running the system ipython (without the module I wanted).

    Poništi
  9. proslijedio/la je Tweet
    4. velj

    The names of 100 prisoners are put in 100 boxes. One at a time, privately, each prisoner opens and closes 50 boxes. They are set free if *everyone* finds their own name. Strategy can be discussed beforehand, but otherwise there's no communication. What's the best strategy? 1/23

    Prikaži ovu nit
    Poništi
  10. proslijedio/la je Tweet
    4. velj

    Elastic Weight Consolidation does indeed help our PE model avoid catastrophic forgetting! A little bit better than L2 regularization. I'm not convinced it's worth it for standard production releases (keep em' simple!), but it's very cool to see stuff like this work!

    Poništi
  11. proslijedio/la je Tweet
    3. velj

    This repo is full of amazing awesomeness. I don't know of anything else like it. Independent refactored carefully tested implementations of modern CNNs

    Poništi
  12. proslijedio/la je Tweet
    , , i još njih 3
    Poništi
  13. proslijedio/la je Tweet
    2. velj

    Performance artist generates virtual traffic jams in Google Maps by pulling a wagon full of smartphones

    Poništi
  14. proslijedio/la je Tweet
    2. velj

    There is a 2016 study "Can we still avoid automatic face detection?" by Cornell University which tests various camouflage techniques, including CV dazzle style makeup, against Facebook's face detection algo and the makeup is not able to fool the algo

    Extract from paper about automatic face detection methods revealing that CV dazzle cannot fool facebook's face detection algorithm
    Poništi
  15. proslijedio/la je Tweet
    2. velj

    "Giving BERT a calculator" Big step in adding computational abilities to language models. Still no understanding but can certainly start to fool you. What happens if you "give BERT" thousand other similar abilities? cc

    Poništi
  16. proslijedio/la je Tweet
    1. velj

    If you need a name for a new malware, threat group operation or hack tool, check this Imgur gallery with 29 posters showing deities, monsters and heroes of different mythologies

    Poništi
  17. proslijedio/la je Tweet
    1. velj

    There is simply no way to safely test if a given cv dazzle pattern will fool the facial recognition algorithm before you go out wearing it.

    Prikaži ovu nit
    Poništi
  18. 1. velj

    Just imagine having to go to your client's HR department and say "Yeah... So turns out we just gave a 'highly recommended' rating to *literally anyone* with a Rothko print in frame, regardless of what they said. Sorry. That's multiple correlation testing for you, huh?"

    Prikaži ovu nit
    Poništi
  19. 1. velj

    Whole new world of adversarial patches, though. "Is this person in front of the camera still Bob?" "FIIK but he's got a weird print on the wall behind him so I bet he's a hard worker!"

    Prikaži ovu nit
    Poništi
  20. 1. velj

    We can't get facial *recognition* to work on people of color with acceptable success rates but yet we're forging ahead into cramming videos of applicants faces through piles of linear algebra to decide if they'd be a good hire? JFC. Enjoy your lawsuits. You deserve them.

    Prikaži ovu nit
    Poništi
  21. proslijedio/la je Tweet
    1. velj

    Facebook for Android crashes in the recent build because I did not give it sufficient permission for location collection

    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:

    ·