\mathfrak{Michael "El Muy Muy" Betancourt}

@betanalpha

Once and future physicist masquerading as a statistician. Reluctant geometer. developer. Support my writing at .

New York, NY
Vrijeme pridruživanja: rujan 2009.

Tweetovi

Blokirali ste korisnika/cu @betanalpha

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

  1. Prikvačeni tweet

    Remember that using Bayes' Theorem doesn't make you a Bayesian. Quantifying uncertainty with probability makes you a Bayesian.

    Poništi
  2. By far the best introduction to probabilistic computation i have seen so far: I also love the juxtaposition of code and graphs. By /ht

    Poništi
  3. prije 22 sata

    Open Access UCL Research: On the Geometric Ergodicity of Hamiltonian Monte Carlo - UCL Discovery

    Poništi
  4. 3. velj

    my roommate calls her significant other her “sig fig” and if that isn’t the most precious thing i don’t know what is

    Poništi
  5. 4. velj

    Can’t wait to see how many people’s supervisors I’ve offended later this week by saying that if you don’t understand when your favourite parameter estimation method fails you shouldn’t be using it

    Poništi
  6. I don't think I love anything in this world more than using Hamiltonian Monte Carlo to track down pathologies in modeling techniques that motivate better priors, experimental designs, and the like.

    Poništi
  7. 4. velj

    I’ll be teaching one of the skills showcase, “Getting started with hidden Markov models”, at I’ll be covering a lot of good stuff! And have lots of advice about how to formulate these models for various data sets.

    Prikaži ovu nit
    Poništi
  8. I doubt this will be particularly impactful but the math was fun--summation by parts!-- and in the process I finally figured out what functional was needed to get the continuous adjoint method to spit out the Jacobian-adjoint product needed for efficient reverse mode autodiff.

    Prikaži ovu nit
    Poništi
  9. New paper from myself, Charles Margossian, and , ! We derive a discrete analogy to the adjoint method for ODEs and use it to work out the gradient of the hidden Markov model marginal likelihood in a new way.

    Prikaži ovu nit
    Poništi
  10. 3. velj

    I used a similarly simple physics lab example to introduce my computational methods students to statistical modeling. This is great!

    Poništi
  11. A course with a view. Looking forward to teaching Bayesian inference this week.

    Poništi
  12. For the morning crowd: I released a new case study yesterday demonstrating principled Bayesian modeling.

    Poništi
  13. 2. velj

    the bishop of orlando is also the bishop of the moon, due to a canon law that says "any newly discovered territory would fall under the bishopric from whence the discovering expedition departed". his is therefore the largest catholic diocese, at over 14,000,000 square miles

    Prikaži ovu nit
    Poništi
  14. 2. velj

    Joaquin Phoenix accepts his Leading Actor award for his performance in

    Poništi
  15. 3. velj

    Yet another fantastic example of how to develop a model from first principles and some trial-and-error:

    Poništi
  16. 3. velj

    What if physics 101 lab were combined with rigorous statistical analysis? What I love about "principled" Bayesian modeling is that it's not a list of a bazillion formulas to memorize (typical of intro stats) but rather a way to combine physical knowledge with data!

    Poništi
  17. 2. velj

    EVERY physicist should plow through this once, and completely, with every integral and every plot. One of the best first-principle-to-wtf-actually-happens demonstrations I ever saw. No 'detections', no 'requires new physics', no 'systematics'. We model, and math is important.

    Poništi
  18. Bonus accompanying soundtrack recommendation: "Falling In Love with the Wolfboy" by the Magnetic Fields. The Twin Peaks theme is another strong option.

    Prikaži ovu nit
    Poništi
  19. As always a huge shout out to my Patreon supporters for making case studies like these possible. If you're interested in supporting my work and getting early access to case studies then consider becoming a supporter, .

    Prikaži ovu nit
    Poništi
  20. Intro physics experiments are more complicated than expected but that's what makes them perfect for demonstrating the utility of Bayesian modeling. In my most recent case study I try to infer gravity from falling ball data and find a few surprises: .

    Prikaži ovu nit
    Poništi
  21. 2. velj
    Odgovor korisniku/ci

    Sure, a calibrated (Probabilistically accurate) but imprecise prediction isn’t good, but neither is a precise but inaccurate one. A good prediction is both. I think Gneiting and others talk about good forecasts aiming to “maximize precision, subject to calibration”. 1/n

    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:

    ·