69 Matching Annotations
  1. Mar 2024
  2. Feb 2024
  3. Oct 2023
  4. Apr 2023
    1. While past work has characterized what kinds of functions ICL can learn (Garg et al., 2022; Laskin et al., 2022) and the distributional properties of pretraining that can elicit in-context learning (Xie et al., 2021; Chan et al., 2022), but how ICL learns these functions has remained unclear. What learning algorithms (if any) are implementable by deep network models? Which algorithms are actually discovered in the course of training? This paper takes first steps toward answering these questions, focusing on a widely used model architecture (the transformer) and an extremely well-understood class of learning problems (linear regression).
    1. This is the space where AI can thrive, tirelessly processing these countless features of every patient I’ve ever treated, and every other patient treated by every other physician, giving us deep, vast insights. AI can help do this eventually, but it will first need to ingest millions of patient data sets that include those many features, the things the patients did (like take a specific medication), and the outcome.

      AI tools yes, not ChatGPT though. More contextualising and specialisation needed. And I'd add the notion that AI might be necessary as temporary fix, on our way to statistics. Its power is in weighing (literally) many more different factors then we could statistically figure out, also because of interdependencies between factors. Once that's done there may well be a path to less blackbox tooling like ML/DL towards logistic regression: https://pubmed.ncbi.nlm.nih.gov/33208887/ [[Machine learning niet beter dan Regressie 20201209145001]]

  5. Jan 2023
    1. I've seen a bunch of people sharing this and repeating the conclusion: that the success is because the CEO loves books t/f you need passionate leaders and... while I think that's true, I don't think that's the conclusion to draw here. The winning strategy wasn't love, it was delegation and local, on the ground, knowledge.

      This win comes from a leader who acknowledges people in the stores know their communities and can see and react faster to sales trends in store... <br /> —Aram Zucker-Scharff (@Chronotope@indieweb.social) https://indieweb.social/@Chronotope/109597430733908319 Dec 29, 2022, 06:27 · Mastodon for Android

      Also heavily at play here in their decentralization of control is regression toward the mean (Galton, 1886) by spreading out buying decisions over a more diverse group which is more likely to reflect the buying population than one or two corporate buyers whose individual bad decisions can destroy a company.

      How is one to balance these sorts of decisions at the center of a company? What role do examples of tastemakers and creatives have in spaces like fashion for this? How about the control exerted by Steve Jobs at Apple in shaping the purchasing decisions of the users vis-a-vis auteur theory? (Or more broadly, how does one retain the idea of a central vision or voice with the creative or business inputs of dozens, hundreds, or thousands of others?)

      How can you balance the regression to the mean with potentially cutting edge internal ideas which may give the company a more competitive edge versus the mean?

  6. Nov 2022
    1. PDF summary by Cochrane for planning a meta-analysis at the protocol stage. Gives guidance on how to anticipate & deal with various types of heterogeneity (clinical, methodological , & statistical). Link to paper

      Covers - ways to assess heterogeneity - courses of action if substantial heterogeneity is found - methods to examine the influence of effect modifiers (either to explore heterogeneity or because there's good reason to suggest specific features of participants/interventions/study types will influence effects of the intervention. - methods include subgroup analyses & meta-regression

  7. Aug 2022
  8. Sep 2021
    1. If you want to find the bug, you can run a mozregression to find what broke it (using 70 as your last known good release and 71 as your bad release).
  9. Jul 2021
  10. Jun 2021
  11. May 2021
    1. Doch die stichhaltigere Erklärung für die Unersättlichkeit des Status- und Machtstrebens liegt in der Regression, d.h. der erlernten Unfähigkeit, im umfassenden Gebrauch der Gesamtheit der eigenen Anlagen Sinn und Erfüllung zu finden, und der daraus resultierenden Verführbarkeit durch die attraktiven Eigenschaften der Macht.“
  12. Apr 2021
  13. Mar 2021
    1. Why separate out red tests from green tests? Because my green tests serve a fundamentally different purpose. They are there to act as a living specification, validating that the behaviors work as expected. Regardless of whether they are implemented in a unit testing framework or an acceptance testing framework, they are in essence acceptance tests because they’re based upon validating behaviors or acceptance criteria rather than implementation details.
    1. Sometimes a change impact analysis is performed to determine an appropriate subset of tests

      Hey, I do that sometimes so I can run a smaller/faster subset of tests. Didn't know it had a fancy name though.

    2. non-regression testing

      That would probably be a better name because you're actually testing/verifying that there hasn't been any regression.

      You're testing for the absence of regression. But I guess testing for one also tests for the other, so it probably doesn't matter. (If something is not true you know it is false, etc.)

    3. Regression testing (rarely non-regression testing[1]) is re-running functional and non-functional tests to ensure that previously developed and tested software still performs after a change.[2] If not, that would be called a regression.
  14. Feb 2021
  15. Oct 2020
  16. Sep 2020
    1. exposure limits are determined by the following equations (NIOSH, 2016):(4)ÂRAL[°C-WBGT]=59.9-14.1log10M<math><mrow is="true"><mi mathvariant="normal" is="true">R</mi><mi mathvariant="normal" is="true">A</mi><mi mathvariant="normal" is="true">L</mi><mo stretchy="false" is="true">[</mo><mi is="true">Â</mi><mi is="true">°</mi><mi mathvariant="normal" is="true">C</mi><mo is="true">-</mo><mi mathvariant="normal" is="true">W</mi><mi mathvariant="normal" is="true">B</mi><mi mathvariant="normal" is="true">G</mi><mi mathvariant="normal" is="true">T</mi><mo stretchy="false" is="true">]</mo><mo is="true">=</mo><mn is="true">59.9</mn><mo is="true">-</mo><mn is="true">14.1</mn><mi mathvariant="normal" is="true">l</mi><mi mathvariant="normal" is="true">o</mi><mi mathvariant="normal" is="true">g</mi><mn is="true">10</mn><mi mathvariant="normal" is="true">M</mi></mrow></math>(5)Â

      regressional analysis of exposure limits

    1. Ip, A., Ahn, J., Zhou, Y., Goy, A. H., Hansen, E., Pecora, A. L., Sinclaire, B. A., Bednarz, U., Marafelias, M., Mathura, S., Sawczuk, I. S., Underwood, J. P., Walker, D. M., Prasad, R., Sweeney, R. L., Ponce, M. G., LaCapra, S., Cunningham, F. J., Calise, A. G., … Goldberg, S. L. (2020). Hydroxychloroquine in the treatment of outpatients with mildly symptomatic COVID-19: A multi-center observational study. MedRxiv, 2020.08.20.20178772. https://doi.org/10.1101/2020.08.20.20178772

  17. Aug 2020
  18. Jul 2020
  19. Jun 2020
  20. May 2020
  21. Jan 2020
    1. An outlier is a data point whose response y does not follow the general trend of the rest of the data. A data point has high leverage if it has "extreme" predictor x values. With a single predictor, an extreme x value is simply one that is particularly high or low. With multiple predictors, extreme x values may be particularly high or low for one or more predictors, or may be "unusual" combinations of predictor values (e.g., with two predictors that are positively correlated, an unusual combination of predictor values might be a high value of one predictor paired with a low value of the other predictor).
    1. As shown in the Residuals vs Fitted plot, there is a megaphone shape, which indicates that non-constant variance is likely to be an issue.
  22. Jul 2019
  23. Jan 2019
    1. There are some environmental elements of the Withdrawal Agreement which our current proposals do not cover, namely those concerning the independent body’s scope to enforce implementation of the “non-regression” clause. We will consider these provisions of the Withdrawal Agreement ahead of publishing the final Bill

      hmmmmm....

    2. The text sets out that, if the protocol is required, the UK and EU will not reduce their respective levels of environmental protection below those in place at the end of the implementation period

      note the 'if' attached to N-R

  24. Sep 2018
    1. 生成模型 vs. 判别模型

      总体来看,如果样本足够多,判别模型的正确率高于生成模型的正确率。

      生成模型和判别模型最大的区别在于,生成模型预先假设了很多东西,比如预先假设数据来自高斯,伯努利,符合朴素贝叶斯等等,相当于预先假设了 Hypothesis 函数集,只有在此基础上才有可能求出这个概率分布的参数。

      生成模型,进行了大量脑补。脑补听起来并不是一件好事,但是当你的数据量太小的时候,则必须要求你的模型具备一定的脑补能力。

      判别模型非常依赖样本,他就是很传统,死板,而生成模型比较有想象力,可以“想象”出不存在于当前样本集中的样本,所以他不那么依赖样本。

      关于 想象出不能存在于当前样本集的样本 ,见本课程 40:00 老师举例。

      生成模型在如下情形比判别模型好:

      1. 数据量较小时。
      2. 数据是noisy,标签存在noisy。
      3. 先验概率和类别相关的概率可以统计自不同的来源。

      释疑第三条优点:老师举例,在语音辨识问题中,语音辨识部分虽然是 DNN --- 一个判别模型,但其整体确实一个生成模型,DNN 只是其中一块而已。为什么会这样呢?因为你还是要去算一个先验概率 --- 某一句话被说出来的概率,而获得这个概率并不需要样本一定是声音,只要去网络上爬很多文字对话,就可以估算出这个概率。只有 类别相关的概率 才需要声音和文字pair,才需要判别模型 --- DNN 出马。

  25. Aug 2018
  26. Oct 2017
  27. Jun 2016
    1. The parameter estimate for the first contrast compares the mean of the dependent variable, write, for levels 1 and 2 yielding 11.5417 and is statistically significant (p<.000). The t-value associated with this test is 3.5122.  The results of the second contrast, comparing the mean of write for levels 1 and 3. The expected difference in variable write between group 1 and 3 is 1.7417 and  is not statistically significant (t = 0.6374, p = .5246), while the third contrast is statistically significant. Notice that the intercept corresponds to the cell mean for race = Hispanic group.

      Interpreting the reference group in dummy coding.

  28. Nov 2015
  29. Aug 2015
    1. Quality (DRG) 1.2675 0.61

      It seems there is no statistically significant differences in the Quality measured between the two aggregation levels.

    2. total offive variables in the frontierestimate: three input types in dollars, output in number of patient-days, andquality in estimated HSMR values.

      So this is the model translated as:

      number of patient-days(by hospital*) = HSMR value(ratio) + capital prices($) + labor($) + materials($)

      • all the co-variants are also accounted by Hospital
    3. bootstrap-adjustedTobit regression as specified by Simar and Wilson (2007

      *interesting reference! on "bootstrap-adjusted Tobit regression!

  30. Jan 2015
    1. Logistic regression, also called a logit model, is used to model dichotomous outcome variables. In the logit model the log odds of the outcome is modeled as a linear combination of the predictor variables.