其他2016cs.toronto.edu
计算机科学成绩并非双峰分布的证据
Evidence That Computer Science Grades Are Not Bimodal
进阶 · 约 35 分钟读完前置知识：基础统计学、分布与直方图、CS 本科课程背景
导读
这篇论文检验“计算机科学课程成绩呈双峰分布”这一常见说法：作者收集了 UBC 1996–2013 年所有本科 CS 课程的最终成绩（778 个教学班、30214 个成绩），并辅以对教师判断直方图的调查。总体证据不支持 CS 成绩普遍双峰，但评论指出部分 100-level 入门课确实更常出现多峰分布。
HN 关注点：是否推翻“CS 成绩双峰”的刻板印象，以及论文的成绩统计和教师调查方法是否站得住脚。
HN 评论精选
这几条评论里：质疑 2 补充 1 亲历 1
亲历CS 教学法仍不成熟，很多人靠自学而非课堂教学学会。
我和许多研究生同学一起尝试给没有 CS 背景的新生讲授计算机科学概念，我的结论是：我们根本不知道如何教 CS。我怀疑很多人学会 CS 是克服了教学，而不是因为教学。想想人类花了多久才学会教所有人阅读：在有硬科学证据表明自然拼读（phonics）是让最后一部分学生学会阅读所需的关键之后，我们又花了几十年才推翻旧的整词阅读法。我们可能还要几十年才能理解如何把 CS 教给所有人，更别说落实这些方法了。
hyperion2010查看英文原文
Having worked with many of my fellow graduate students to try to teach CS concepts to incoming students without any CS background my conclusion here is that we simply don't know how to teach computer science. My suspicion is that many people who learn CS usually do it inspite of attempts to teach them, not because of instruction. Think about how long it took humans to figure out how to teach reading to _everyone_. We had to break the old whole reading method and switch over to phonics, and that was decades after we had the hard scientific evidence that phonics was the missing perspective needed to get that last percentage of students to read. We are probably decades away from understanding how to teach CS to everyone, let alone getting those techniques implemented.
质疑质疑论文对 Padavic 等文献的转述不准确。
这段和文章主题基本无关，但下面这段引起我注意：Padavic 等人发现商业领域的“工作—家庭”叙事是一种社会防御：人们会说女性因“家庭”离开职场，尽管大量证据表明女性离职是因为薪酬或晋升机会不足，尤其是看到男同事被提拔时。“工作—家庭”叙事比直面职场性别歧视更容易被接受，所以继续流传。我试图追踪 Patitsas 等人引用的 Padavic 等工作并找到了原文；但 Padavic 等人的实际论点与 Patitsas 等人的概括不同：他们似乎是在论证“工作—家庭”叙事是对公司里两性都超长工作时间这一不快事实的社会防御，而不是针对性别歧视；他们似乎也完全没有讨论薪酬或晋升中的性别歧视。Patitsas 等人是不是引用有误？还是我漏看了什么……
a_puppy查看英文原文
This is mostly unrelated to the main topic of the article, but this paragraph caught my eye: > For example, Padavic et al. [20] found that the "work-family" narrative in business is an example of a social defense: people will say that women leave the workplace because of "family", despite the large amount of evidence that women leave their jobs because of inadequate pay or opportunities for advancement [20], particularly when they see male co-workers promoted ahead of them. The "work-family" narrative is a more palatable explanation rather than to confront sexual discrimination in the workplace, and so the narrative continues. > > [20] I. Padavic and R. J. Ely. The work-family narrative as a social defense, 2013. I tried to track down the work by Padavic et al. that Patitsas et al. are citing, and found this: http://www.hbs.edu/faculty/conferences/2013-w50-research-symposium/Documents/Gender_and_work_web_update2015.pdf However, this article by Padavic et al. is actually saying something different from the way that Patitsas et al. summarized it. Padavic et al. appear to actually be arguing that the work-family narrative is a social defense against the uncomfortable truth of excessively long working hours for both genders at the corporation, not against any uncomfortable truth about sexual discrimination. Also, Padavic et al. don't seem to discuss sexual discrimination in wages or promotions at all. Did Patitsas et al. make an error in citing their sources? Am I missing something?
补充多峰分布集中在 100 级入门课，需单独细看。
“CS 成绩是双峰还是单峰？为检验这一点，我们获取了 UBC 1996 到 2013 年每个本科 CS 课程的最终成绩分布。这涵盖 778 个不同的教学班，共 30214 个最终成绩（平均班级规模 75 人）。” 我对双峰现象的理解是：如果它存在，主要应出现在较早的课程——后续课程只包含那些之前课程学得好或至少通过的人。啊，果然……“在 45 个多峰分布的班级中，16 个是 100 级（35%），5 个是 200 级（11%），12 个是 300 级（27%），12 个是 400 级（27%）。作为对照，全部 778 个班级中，171 个是 100 级（22%），165 个是 200 级（21%），243 个是 300 级（31%），199 个是 400 级（26%）。” 我们是不是该更仔细看看那些 100 级课程？
mcguire查看英文原文
"Are CS grades bimodal, or unimodal? To test this, we ac- quired the final grades distributions for every undergraduate CS class at the University of British Columbia (UBC), from 1996 to 2013. This represents 778 different lecture sections, containing a total of 30,214 final grades (average class size: 75)." My understanding of the bimodal situation, if it exists, is that it primarily applies to earlier classes---later classes only include those who did well, or at least passed, the previous classes. Ah, yes... "Of the 45 classes which were multimodal, 16 were 100- level classes (35%), 5 were 200-level (11%), 12 were 300-level (27%), and 12 were 400-level (27%). For comparison, in the full set of 778 classes, 171 were 100-level (22%), 165 were 200-level (21%), 243 were 300-level (31%), and 199 were 400-level (26%)." How about we take a closer look at those 100 level classes, hmmm?
质疑调查所用直方图因截断和挑选产生虚假双峰，方法有偏。
我认为论文调查部分的整个前提很有问题。其设想是“给教授看一些我们知道是正态的直方图，看他们是否会看出并不存在的双峰”。问题是，他们并没有给教授看 6 个随机正态分布直方图。首先，他们展示的甚至不是真正的正态分布，而是截断在 100 分的正态分布；这会在 100 分处形成一个质点，从而使其变成双峰分布！其次，他们也不是随机给 6 个截断正态直方图，而是从一组随机直方图中挑了 6 个，其中 4 个看起来像双峰。参与者看到的不是不存在的双峰，而是由直方图挑选方法刻意制造出来的双峰！
scott00查看英文原文
The whole premise of the survey part of this paper strikes me as very flawed. The idea was "show some professors histograms we know are normal, and see if they see bimodality that isn't there". The problem is, they didn't show them 6 random histograms of the normal distribution. For one, they didn't actually show them normal distributions at all; they showed them normal distributions capped at 100. That's going to give you a point mass at 100... making it a bimodal distribution! Second, they didn't give 6 random examples of a capped normal histogram; they gave them 6 examples chosen from a random set of histograms chosen in such a way that 4 of the 6 look bimodal. The survey participants didn't see bimodality that wasn't there, they saw bimodality that was purposefully generated by the histogram selection methodology!
相关资源
从这篇论文所有 HN 讨论帖的全部评论里整理出来的链接。
历次讨论
- 2016Hacker News 讨论帖146 分 · 140 评
同领域还值得读
每个好的系统调节器都必须是该系统的一个模型 (1970)
Every Good Regulator of a System Must Be a Model of That System (1970)
这是控制论经典论文（Conant 与 Ashby，1970）：它试图论证，任何“好的”调节器都必须是其所调节系统的模型——即调节器与被控系统之间必须存在同构或模型关系。
其他2018 · 202 分 · 59 评进阶 · 约 45 分钟编辑精选复杂系统如何失效
How complex systems fail (2002)
总结了复杂系统如何失败的若干条经验法则：系统总是在多重故障的边缘运行，事故是多重因素叠加的结果，操作者实际上同时承担着维持生产与防范故障的双重角色，而事后归因往往把事故简化为某个…
其他2017 · 154 分 · 19 评入门 · 约 30 分钟编辑精选如何阅读论文
How to Read a Paper (2016)
介绍如何高效阅读学术论文，提出“三遍阅读法”：先快速浏览摘要、结论和图表把握概览，再理解主要内容，最后深入细节。
其他2017 · 715 分 · 80 评入门 · 约 10 分钟公元 1021 年欧洲人已在美洲存在的证据
Evidence for European presence in the Americas in AD 1021
这篇 Nature 论文用树木年轮定年，并结合公元 993 年宇宙射线事件标记，把纽芬兰 L'Anse aux Meadows 的 Norse 活动精确到公元 1021 年，为欧…
其他2021 · 655 分 · 391 评进阶 · 约 45 分钟一次通古斯级空爆摧毁了约旦河谷青铜时代城市 Tall el-Hammam
A Tunguska size burst destroyed Tall el-Hammam, Bronze Age city in Jordan Valley
研究约旦河谷青铜时代城市 Tall el-Hammam 被毁的原因。
其他2021 · 653 分 · 342 评入门 · 约 45 分钟