Smerity

@Smerity

gcc startup.c -o . Focused on machine learning & society. Previously Research via . '14, '11. AU in SF.

San Francisco, CA
smerity.com
Geregistreerd in juli 2008
664 foto's en video's Foto's en video's

Tweets

Je hebt @Smerity geblokkeerd

Weet je zeker dat je deze Tweets wilt bekijken? @Smerity wordt niet gedeblokkeerd door Tweets te bekijken.

  1. Vastgemaakte Tweet
    23 mrt.

    Simplicity is still competitive! Standard LSTM/QRNN models can achieve state-of-the-art results on character (PTB, enwik8) + word level (WikiText-103) language modeling datasets in 12-48 hours w/ a single GPU. Joint work w/ &

    Character-level Penn Treebank results
    Character-level enwik8 results
    WikiText-103 results
    Analysis of hyper-parameter sensitivity
    15 antwoorden 190 retweets 560 vind-ik-leuks
    Deze collectie tonen
    Deze collectie tonen
    Ongedaan maken
  2. 1 uur geleden

    The format is concise, so you can't wax poetic or reply to multiple thematic threads at once, and your audience is whoever subscribes to you, which is anyone who is interested. There are also essentially no abstraction layers between those communicating. No legal team. No PR.

    1 retweet 7 vind-ik-leuks
    Deze collectie tonen
    Deze collectie tonen
    Ongedaan maken
  3. 1 uur geleden

    Whenever I say (paraphrased) "Twitter still houses the best community for even though that theoretically and practically makes no sense", this is exactly what I mean. Ever more direct connections between those working in, reporting on, or learning about a field.

    Great new study about science outreach via Twitter: Initially, scientists mostly tweet to each other. But after accumulating about 1000 followers, scientists reach an increasing number of journalists, policy makers, and other members of the public.
    2 antwoorden 5 retweets 41 vind-ik-leuks
    Deze collectie tonen
    Deze collectie tonen
    Ongedaan maken
  4. heeft geretweet
    2 uur geleden

    Having the right people in a team is more important than having access to compute resources.

    6 antwoorden 26 retweets 103 vind-ik-leuks
    Deze collectie tonen
    Deze collectie tonen
    Ongedaan maken
  5. 3 uur geleden

    There is a worrying grey area in what activity and behaviour is tracked and used to inform predictive algorithms when you have no choice or transparency. As opposed to viewing history I usually can't tell a given service to forget about my mouse click or content page view.

    2 antwoorden 1 retweet 6 vind-ik-leuks
    Deze collectie tonen
    Deze collectie tonen
    Ongedaan maken
  6. 3 uur geleden

    Real thought: I was flicking through <online video service> and then hesitated about hovering on an obviously pseudoscience series that I was curious about. Why? I was concerned whether it would register this as genuine interest rather than irate intellectual curiousity.

    2 antwoorden 2 retweets 16 vind-ik-leuks
    Deze collectie tonen
    Deze collectie tonen
    Ongedaan maken
  7. 4 uur geleden

    My real pride there is in the definition of dropout having been cited* by 199 of the deep learning community's finest scholars. * If retweets** are endorsements ** If endorsements are citations

    Dropout (Srivastava et al., 2014) may be the first instance of a human curated artisanal regularization technique that entered wide scale use in machine learning. Dropout, simply described, is the concept that if you can learn how to do a task repeatedly whilst drunk, you should be able to do the task even better when sober.
    If you ever need a definition of dropout that is both concise and accurate:
    Deze collectie tonen
    1 antwoord 2 retweets 9 vind-ik-leuks
    Deze collectie tonen
    Deze collectie tonen
    Ongedaan maken
  8. 4 uur geleden

    If you were interested in my (comparably weak) SIGBOVIK deep learning paper, I refer you to my profoundly novel work in "The NUGGET Non-Linear Piecewise Activation":

    ReLU isn't the only piecewise activation function to work well. Introducing the NUGGET Non-Linear Piecewise Activation, the result of a complex zero-sum pricing game over epochs of multi-agent interaction. Independent work in proceedings of SigBovik 2018.
    Deze collectie tonen
    1 antwoord 3 vind-ik-leuks
    Deze collectie tonen
    Deze collectie tonen
    Ongedaan maken
  9. 4 uur geleden

    Authors would be far more motivated to do in-depth analysis if they got to act like a sports commentator narrating Puppy Bowl XIII whilst assessing the vast stupidity that their algorithms are capable of ^_^

    Figure 11: Pac-Man showing utter disregard for the ghosts’s personal space. This occurs around frame 6700 of the output movie. Pac-Man slips into the space between Blinky and Inky to touch the wall, comes out unharmed, then momentarily teases Clyde before escaping to vibrate some more in empty corridors.
    Figure 9: Mario bounces off one Goomba up into the feet of another. Not only doesn’t he have enough velocity to reach the platform, but he’s about to hit that Goomba from below. Believe it or not, this ends well: playfun is happy to exploit bugs in the game. (cut off due to text limits)
    Figure 13: Bub navigates this game surprisingly well. Once he’s on his last life, he becomes careful and pauses the game when things are looking grim—here pausing for about a thousand frames, burning through the futures until one randomly comes along that looks good. He then unpauses and executes that good future, killing three of these monsters in short order.
    1 antwoord 2 retweets 5 vind-ik-leuks
    Deze collectie tonen
    Deze collectie tonen
    Ongedaan maken
  10. 4 uur geleden

    If you want a fun read, I highly recommend the paper, which seems more thorough than many deep learning papers 😅 It's also so much more honest than academic papers - you don't have to pretend to be smart! All DL papers need "Parameter Alert!" for example.

    Parameter Alert! This one of the first places where we have some arbitrary constants, which are the enemy of elegance. On a scale of ballpark to obsessively overfit, these constants are a 2; I basically looked at some graphs while developing the objective function learning part of the code to decide whether they were “good enough” and didn’t tune them after starting to observe actual performance.
    2 antwoorden 2 retweets 6 vind-ik-leuks
    Deze collectie tonen
    Deze collectie tonen
    Ongedaan maken
  11. 4 uur geleden

    Throwback to: 's SIGBOVIK paper which showed a machine imbued with time travel fails to beat Tetris and instead decides to pause forever. Truly a wise algorithm. Video: (Bias: I am also a SIGBOVIK author. Such prestige, much wow. SIGBOVIK >>> NIPS.)

    Figure 16: Would you hire this construction company? Death is imminent, so playfun pauses the game shortly after this and then doesn’t unpause it.
    1 antwoord 7 vind-ik-leuks
    Deze collectie tonen
    Deze collectie tonen
    Ongedaan maken
  12. heeft geretweet
    24 uur geleden

    "World War III is a guerrilla information war with no division between military and civilian participation." Marshall McLuhan (1970)

    29 antwoorden 870 retweets 1.899 vind-ik-leuks
    Ongedaan maken
  13. heeft geretweet
    24 uur geleden

    So, 500px, which positioned itself as a Flickr alternative, has announced they're deleting/removing access to all Creative Commons photos on their site. Like, tomorrow. NO warning.

    14 antwoorden 183 retweets 126 vind-ik-leuks
    Deze collectie tonen
    Deze collectie tonen
    Ongedaan maken
  14. 28 jun.

    First figure from our paper: how the LSTM with a twist allows for the equivalent speed of a plain convnet by running efficiently in parallel on GPUs, like image processing convents.

    4 antwoorden 9 retweets 40 vind-ik-leuks
    Deze collectie tonen
    Deze collectie tonen
    Ongedaan maken
  15. 28 jun.

    Actually, yes, the QRNN has all those features. This is the Hacker News equivalent of seeing "Reviewer confidence: 5" after realizing from their comments the reviewer actually hasn't read the paper. (This is what I get for reading the Hacker News comments on Keras vs PyTorch ...)

    https://news.ycombinator.com/item?id=17418155
    3 antwoorden 4 retweets 50 vind-ik-leuks
    Deze collectie tonen
    Deze collectie tonen
    Ongedaan maken
  16. 28 jun.

    The future is where you may lose 1 out of every 333 of your customers due to a suboptimal experience caused by a psychological experiment where neither the content creator nor viewer gave consent or received notice. Isn't it nice to be a sentient cog in the big data machine?

    Als antwoord op @AlarconGareca
    We are running a small experiment where 0.3% of viewers will see an auto-generated thumbnail, instead of your custom thumbnail. We are not removing the ability to create your custom thumbnail, but we hope to gain insights on auto-generated thumbnails for the future.
    5 antwoorden 21 retweets 125 vind-ik-leuks
    Ongedaan maken
  17. heeft geretweet
    26 jun.

    Interesting thread on the OpenAI bot that learned how to play a complex 5-vs-5 video game through 120 years' worth of simulation, every day, for months. Its secret weapon: Not having any of the humans' emotional reactions or selfish instincts. The ultimate team player

    The bots knew when to take the fall to set their team up, when to waive their local reward for the team's global reward, and had no sense of a hero complex (I'm going to attack as I have a time limited Double Damage rune even though my team isn't ready!).
    Deze collectie tonen
    1 antwoord 10 retweets 22 vind-ik-leuks
    Ongedaan maken
  18. heeft geretweet
    26 jun.

    This isn’t algorithmic bias, like many systems it’s the hiding of human opinion and bias behind the veneer of an algorithm.

    This is wild. All of us tech researchers are sitting here warning about the small, subtle ways algorithmic bias can support fascism — and meanwhile ICE has essentially created a button that prints “detain immigrant” when pressed
    Deze collectie tonen
    3 antwoorden 24 retweets 41 vind-ik-leuks
    Ongedaan maken
  19. heeft geretweet
    26 jun.

    Bizarre false equivalence by Work. Of course there’s espionage of/repurposing of Google work *in* China by China’s gov’t, as there is for work outside China, but that’s different from working *for* China. Maven was about working *for* US.

    Fmr DepSecDef Bob Work slams Google, says Project Maven withdrawal risks US lives, notes Google’s continuing work for China:
    4 antwoorden 3 retweets 12 vind-ik-leuks
    Deze collectie tonen
    Deze collectie tonen
    Ongedaan maken
  20. heeft geretweet
    26 jun.

    Excited to release new and updated text-to-SQL datasets + observations of limitations in current evaluations (and solutions)! and A great collaboration with Catherine Finegan-Dollak, and more from

    1 antwoord 22 retweets 79 vind-ik-leuks
    Ongedaan maken
  21. heeft geretweet
    25 jun.

    Looking forward to testifying tomorrow on AGI. It's important for Silicon Valley to have an open dialogue with policymakers about this incredibly powerful potential technology.

    OpenAI co-founder will be testifying on Capitol Hill on Artificial General Intelligence, Tuesday at 10:30a ET. Hearing will be livestreamed:
    10 antwoorden 40 retweets 260 vind-ik-leuks
    Ongedaan maken

@Smerity heeft nog niet getweet.

Het laden lijkt wat langer te duren.

Twitter is mogelijk overbelast of ondervindt een tijdelijke onderbreking. Probeer het opnieuw of bekijk de Twitter-status voor meer informatie.

    Je bent misschien ook geïnteresseerd in

    ·