Johan Carlin

@johancarlin

Brain scientist.

Cambridge, UK
Vrijeme pridruživanja: prosinac 2011.

Tweetovi

Blokirali ste korisnika/cu @johancarlin

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

  1. proslijedio/la je Tweet
    3. velj

    I made an illustration for my class of why you need separate validation and training sets and I'm quite happy with it. Has anyone illustrated this similarly before?

    Prikaži ovu nit
    Poništi
  2. 31. sij

    A little blog post on noise ceilings. If you've ever wondered about the relation between the noise ceilings you come across in the encoding model literature and the ones used in RSA, this might help:

    Poništi
  3. proslijedio/la je Tweet
    29. sij

    Since I first read it, I've never stopped thinking about "The Natural Selection of Bad Science" The paper demonstrates better than any other how science reform is more about restructuring incentives than winning logical arguments

    Poništi
  4. 29. sij

    And if you hit <tab> while composing a reply, the first selected button below is discard, not send. Helps cut down on emails!

    Prikaži ovu nit
    Poništi
  5. proslijedio/la je Tweet
    29. sij

    A guideline for a research using Bayesian Statistics: a good how-to for people who would like to Bayes factor instead of ordinary testing, but I am wondering why they do not refer to posterior predictive distribution (not 'posterior distribution').

    Prikaži ovu nit
    Poništi
  6. proslijedio/la je Tweet
    29. sij

    Was thrilled to take the 💯⭐️⭐️⭐️⭐️⭐️ workshop “Deep Learning with Keras and TensorFlow in R” from Observation: So much term rebranding in deep learning. I have a section of notes of the form: {ye olde stats term} -> {fancy-pants deep learning term}

    Prikaži ovu nit
    Poništi
  7. 22. sij

    A tutorial on how power and significance thresholds influence reproducibility. Why did Ioannidis claim that most published findings are false? Would a stricter criterion for significance help? Spent way too long on this. Hope it's useful for someone

    Poništi
  8. 20. sij

    Using boundary-based registration to improve motion correction in highly distorted 7T fMRI data. Our latest is now in press (OA):

    Poništi
  9. proslijedio/la je Tweet
    19. sij

    Some people think we don’t *really* know something until a lucky person convinces 3 others they should be the one to write it down in a big book that gets added on to 4 times/year, which nobody reads but all of us pay for, so they can get a better job title but no more money.

    Poništi
  10. 7. sij

    £10 breakfast at Stockholm airport. Possibly the city with the greatest city-airport discrepancy in food quality (in town you get gouged too, but these days you probably get a decent sandwich).

    Poništi
  11. proslijedio/la je Tweet
    2. sij

    Seeing twitter debate about the replicability of the new DeepMind study of ML for breast cancer screening vs. the NYU study from October: Did some light digging in to compare/contrast… 1/n

    Prikaži ovu nit
    Poništi
  12. proslijedio/la je Tweet
    25. pro 2019.

    There are two facts which make correlations controlling for confounds uninformative about many (but not all) causal effects: 1) the R^2 of the mechanisms we understand is low, 2) our uncertainty about not well-understood mechanisms should be high. (1 / about 13-15)

    Prikaži ovu nit
    Poništi
  13. 20. pro 2019.

    That's all I have. Please check it out. PRs welcome! Goodnight.

    Prikaži ovu nit
    Poništi
  14. 20. pro 2019.

    New in this release: 1. bleeding edge package versions across the board thanks to a better automated build process with make. 2. Shell script wrappers to handle non-conda dependencies (eg, fsl, freesurfer). More work needed here to generalise past cbu (maybe container?).

    Prikaži ovu nit
    Poništi
  15. 20. pro 2019.

    Non-technical users enjoy analysing data instead of dealing with packages and environments. But the real payoff for me is in reproducibility. When those users share code, other people have a decent chance of a) running it and b) getting the same results.

    Prikaži ovu nit
    Poništi
  16. 20. pro 2019.

    Each Neuroconda release is intended to be deployed as its own environment. Users just activate neuroconda_2_0 and get on with analysis. Users report the neuroconda release used in outputs. This provides a fairly complete description of their environment in a few words.

    Prikaži ovu nit
    Poništi
  17. 20. pro 2019.

    I work in the methods group at . Our users are interested in reproducibility and open science, but struggle with technical demands. Maintaining a consistent environment and reporting it accurately with your results is *hard* if you're not already a developer.

    Prikaži ovu nit
    Poništi
  18. 20. pro 2019.

    Neuroconda 2.0 is out: . A complete environment for neuroscience in python, R, and more. Why? Reproducibility. Thread:

    Prikaži ovu nit
    Poništi
  19. 20. pro 2019.

    So that's what all the TPUs are for! "if you do a lot of runs, you’re able to report only the best ones. Results from the people with more computing power to do more runs will look better.”

    Poništi
  20. 26. stu 2019.

    'cargo cult scientists “follow all the apparent precepts and forms.” The problem is mostly not disregard for epistemic norms; it is that the norms themselves are inadequate.'

    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:

    ·