Shital Shah

@sytelus

Research Engineer, Microsoft Research AI. Created AirSim & TensorWatch. A program trying to understand what it's computing.

Redmond, WA
Vrijeme pridruživanja: srpanj 2007.

Tweetovi

Blokirali ste korisnika/cu @sytelus

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

  1. prije 15 sati

    Coronavirus cases were first reported on Dec 31, first death on Jan 11 and first paper was on Jan 23! During past 12 days there have been 30 more papers. Virus DNA already sequenced, similarity with SARS+HIV DNAs identified and potential drugs predicted!!

    Poništi
  2. 2. velj

    Politics is a business where the product is the country. Let's never forget that politicians are just businessmen. It's irrelevant which side of the aisle you are.

    Poništi
  3. proslijedio/la je Tweet
    1. velj

    What (unreleased) Tesla FSD Autopilot sees Autonomy/ electrification unrelated tech bring mutual boost Autonomy largest part of coming disruption full build of Autopilot: 48 neural networks take 70,000 GPU hours to train

    Poništi
  4. proslijedio/la je Tweet
    31. sij
    Odgovor korisniku/ci

    Right now we use about a million atoms per stored magnetic bit. We have some way to go.

    Poništi
  5. 28. sij

    Most people don’t get the difference between a millionaire vs billionaire. If you earn a dollar every second, it will take you a month to be millionaire but it will take you 30 years to be a billionaire. Billion is too big for evolution to inscribe intuition for it.

    Poništi
  6. 23. sij

    It seems that every effort to eliminate one sgd hyperparamter eventually has produced more hyperparameters. Learning rate is like Hydra’s head in deep learning.

    Poništi
  7. 22. sij

    Why do we see a weight decay values like 3x10^-4? Why not 5 or 6 or some other digit? Because this is exponent bisection of 10^-3 and 10^-4, i.e., 10^-3.5 which is 3x10^-4.

    Poništi
  8. proslijedio/la je Tweet
    1. tra 2019.

    If you're bootstrapping yourself into deep learning research, here’s what I would do: 1. FastAI (3m) 2. Personal projects/reproduce papers/consulting (3 - 12m) 3. Flashcard the Deep Learning Book (4-6m) 4. Flashcard ~100 papers in a niche (2m) 5. Publish your first paper (6m)

    Poništi
  9. proslijedio/la je Tweet
    6. lip 2019.

    Interesting AI research areas that require little compute: - Mathematical reasoning - Novel activation functions - Models on 1-8 bits - Working memory - Micro/macro attention - 2nd order optimizations - Model visualizations - Adaptive computation time - Neural ODE What else?

    Poništi
  10. 17. sij

    How does it feel like to accidentally format the USB drive with $3.14M worth of bitcoins? Same as accidentally cashing out the $3.14 check from Donald Knuth.

    Poništi
  11. 16. sij

    Saw new hard to say phrase: Co-opetition - Cooperation disguised as competition. Example: Cable companies are not competing with each other, they are in coopetiting.

    Poništi
  12. 15. sij

    One project is Life-long Kindergarten project . There is a lot needs to be done. Current academic format is towards punishing as opposed to encouraging, competing as opposed to experiencing, be done with fast as opposed explore longer. How would future look like? 8/8

    Prikaži ovu nit
    Poništi
  13. 15. sij

    Many, if not everyone, would then be enrolled in a formal study for most of their lives. Surfing from one micro/mini/full degree to another. Learning shouldn't be just first ~20 years just so we can find a job but a life long journey towards satisfying our own curiosity. 7/n

    Prikaži ovu nit
    Poništi
  14. 15. sij

    I can imagine the future, perhaps 100 yrs down the line, where our collective bandwidth frees up, may be just 3 day work week. Our actions then would be purely be driven by curiosity to know ourselves and nature as opposed to finding job, make money and satisfy basic needs. 6/n

    Prikaži ovu nit
    Poništi
  15. 15. sij

    Making these a formal studies is much better than MOOCs or browsing YouTube videos because it induces structure, interaction with experts, guarantees on what you would learn, adds motivation, collaboration with other students, check on time, tests, measurements, improvement. 5/n

    Prikaži ovu nit
    Poništi
  16. 15. sij

    And this is not even including languages, various sub-fields in CS, ML, robotics, computer vision. Why shouldn't we always be learning something new and interesting? All this knowledge and cool stuff accumulated over thousands of years standing right in front of us! 4/n

    Prikaži ovu nit
    Poništi
  17. 15. sij

    geology, paleontology, botany, organic chemistry, molecular biology, protein synthesis, genetics, quantum computing, industrial mfg, supply chain, macroeconomics, glass blowing, woodworking, pottery science, telescope making, chip design, metallurgy, neuroscience,... 3/n

    Prikaži ovu nit
    Poništi
  18. 15. sij

    I have wondered: Why shouldn't everyone always be enrolled in some studies, accumulating degrees all of their lives? There are so many interesting things to know and learn! I'd like micro/mini-degrees in cooking, gardening, nutrition, astronomy, relativity, music theory, ... 2/n

    Prikaži ovu nit
    Poništi
  19. 15. sij

    This is a very interesting idea: . You get an accredited 4 year CS degree from a legit university while working at a company and getting paid. Students spend 25hr/wk at the company and rest in studies. This arrangement can make possible one cool thing.. 1/n

    Prikaži ovu nit
    Poništi
  20. proslijedio/la je Tweet
    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:

    ·