HackerTube English?
表示、直觉与反馈 →

证据来了怎样改判断,没来之前怎样选问题?

一次检测和一次猜词看似无关,却都要从候选集合中筛掉不符合观察的部分。区别是:你已经看到了答案,还是正要决定问什么?

本文的问题 · 12
  1. 先把不符合证据的人移开
  2. 还没看到颜色时,为什么要算另一种量?
  3. 同一束光、同一段声音,判断为何会变?
  4. 一个没出现的信号,与许多相同的分子
  5. 十个正面,是哪一次尝试的十个?
  6. 额定功率和烧开水的秒数之间漏算了什么?
  7. 看到 4500 瓦,可以算出下一张电费单吗?
  8. 节省的能源到底在边界哪一侧?
  9. 机器省下的功耗,是否转成了使用者的工作?
  10. 把一小时移到傍晚,真正改变了哪些负荷?
  11. 账本有盈余,改变制度就能省下同样一笔钱吗?
  12. 结果出来前,什么会让你改主意?

先把不符合证据的人移开

想象一千张卡片。假设其中 10 张代表带有某种情况的人;检测把这 10 张里的 9 张标红,同时也把另外 990 张里的 89 张标红。拿到一张红卡时,你在 98 张红卡里找那 9 张,不是在 10 张有情况的卡里找它们。9/10 和 9/98 都算对了,却回答了相反方向的问题。这些是视频为演示设定的数字,不是某项检测的实际性能。

原文与时间戳 · 1
And then suppose that among the remainder without cancer, 89 get false positives, and 901 correctly get negative results.
The medical test paradox, and redesigning Bayes' rule · ▶ 1:37–1:44

另一讲把抽卡变成面积:先验比例决定各块原本有多大,观察到描述后,只在符合描述的区域里重新比较。两种讲法可以互相校验:如果你画出的剩余区域对应不到所有红卡,分母就漏了一群人。

原文与时间戳 · 1
In the standard notation, this vertical bar means given that, as in we're restricting our view only to the possibilities where the evidence holds.
Bayes theorem, the geometry of changing beliefs · ▶ 5:10–5:18
把一千张卡片真正摆出来
标红未标红
有该情况91
无该情况89901
自测:如果有人交给你一张红卡,先看哪一列?

只看标红列,分母是 9 + 89。若问有该情况的人中多少被标红,才读第一行,分母是 9 + 1。

还没看到颜色时,为什么要算另一种量?

玩 Wordle 时,猜之前还不知道哪组颜色会亮。不能挑最幸运的一组反馈来评价这个词,而要看每一种反馈可能排除多少不确定性,并按它出现的概率求平均。收到颜色后的更新,才对应上一节「只看剩下的红卡」。前者是在设计问题,后者是在解释答案。

原文与时间戳 · 1
The way we measure the quality of this guess will be to take the expected value of this information.
Solving Wordle using information theory · ▶ 10:43–10:48

一个容易手算的反例:四个候选词中,一个占 70% 的先验概率,其余三个各占 10%。一种猜法能把它们按数量分成两组各两词,可两组的概率只能是 80% 和 20%;如果改问「是不是那个最常见的词」,数量变成一对三,概率反而是更接近对半的 70% 和 30%。在这两个是非问题里,后一个的预期信息更多。视频里甚至有一个含 32 个生僻词的分组,与另一个仅有 8 个词的分组差不多可能。用这种方法挑下一问,还得说清你优化的是预期信息,还是尽快猜中答案。

原文与时间戳 · 1
So a quarter as many matches, but it's about as likely.
Solving Wordle using information theory · ▶ 23:06–23:09

同一束光、同一段声音,判断为何会变?

Lotto 的色块进入眼睛的光一样,但周围看起来一块在阴影里,另一块在亮处。Seth 播完语音线索后重播失真的声音,听者能突然听出先前听不到的话。两例分别改变了对照明和对声源的预期。Hoffman 的桌面图标比喻再多问一步:一个表示有助行动,能否证明内部图景在形状上也像对象?演示能检验的是具体知觉会怎样改变;Hoffman 从进化模型引出的普遍论点还需额外论证。

原文与时间戳 · 3
Because if the tile in that shadow were in fact in shadow, and reflecting the same amount of light to your eye as the one outside the shadow, it would have to be more reflective -- just the laws of physics.
Optical illusions show how we see | Beau Lotto · ▶ 7:29–7:40
All that's changed is your brain's best guess of the causes of that sensory information. And that changes what you consciously hear.
Your Brain Hallucinates Your Conscious Reality | Anil Seth | TED · ▶ 7:00–7:07
Anyone who thought that misinterprets the purpose of the interface. It's not there to show you the reality of the computer. In fact, it's there to hide that reality.
Do we see reality as it is? | Donald Hoffman | TED · ▶ 12:25–12:35

一个没出现的信号,与许多相同的分子

Loeb 描述 ʻOumuamua 时,把没有观测到预期的彗尾列为需要解释的现象。这能挑战一种具体的彗星模型,但还不能单独选出人工起源。Cronin 从另一种观察入手,既数结构的组装量,也数同一种分子出现多少份。比较两段时把纸分成两栏:一个解释预期看到什么,替代解释也能产生什么。无论是没看到一个信号,还是看到了许多复制品,都得先知道两种解释的预期是否真的不同。

原文与时间戳 · 4
there was no cometary tail
Avi Loeb: Aliens, Black Holes, and the Mystery of the Oumuamua | Lex Fridman Podcast #154 · ▶ 23:32–23:35
searched very deeply for carbon-based molecules
Avi Loeb: Aliens, Black Holes, and the Mystery of the Oumuamua | Lex Fridman Podcast #154 · ▶ 23:39–23:44
We count the number of identical molecules
Lee Cronin: Origin of Life, Aliens, Complexity, and Consciousness | Lex Fridman Podcast #269 · ▶ 106:34–106:36
calculate the assembly number
Lee Cronin: Origin of Life, Aliens, Complexity, and Consciousness | Lex Fridman Podcast #269 · ▶ 106:40–106:42

十个正面,是哪一次尝试的十个?

先指定一轮,得到十个正面的概率很小。但如果看到反面就结束该轮,反复投到成功才剪出镜头,视频出现十个正面所回答的问题就变了。讲者把失败轮次也讲清楚,这让人能重建试验规则。遇到只保留成功结果的数据集,也要问失败的尝试保存在哪里。

原文与时间戳 · 2
it's a half to the power of 10
Flipping 10 heads in a row: Full video · ▶ 0:24–0:30
stop the trial as soon as I saw a tail
Flipping 10 heads in a row: Full video · ▶ 1:39–1:45

额定功率和烧开水的秒数之间漏算了什么?

用水所需焦耳除以额定瓦数,可以写下一个可核验的时间预测。实测更慢时,把壶体也会吸热、实际功率和标签有差异这两项单独列出来。这和加筋砂的例子一样:把「一座壶」或「一堆沙」当成单一物体容易算,回到能量或力的实际路径才能解释差值。

原文与时间戳 · 3
334,720 joules divided by 1,100 watts is 304 seconds
Why don't Americans use electric kettles? · ▶ 12:01–12:10
energy needed to heat the kettle itself
Why don't Americans use electric kettles? · ▶ 12:14–12:20
The tension in the reinforcement is generating confining pressure in the soil.
Sand Castle Holds Up A Car! - Mechanically Stabilized Earth · ▶ 4:33–4:37

看到 4500 瓦,可以算出下一张电费单吗?

额定功率描述使用能量的速率;要算用电量,还得知道满载运行多久。电热水器给出一个额外陷阱:两个加热元件放在不同高度,功率可以从一个切到另一个。简单把两个额定数相加,就把空间中存在两根元件误写成两者同时满载。要比较两台电器,让它们完成同一任务,再累计一段时间的用电。

原文与时间戳 · 4
slow the rate of fuel flow, the flame shrinks
Power is not energy: why the difference matters · ▶ 3:46–3:54
you might assume the dryer has a huge impact on your energy bill.
Power is not energy: why the difference matters · ▶ 25:41–25:45
two heating elements sticking into the tank at two different heights.
A simple water heater is more clever than it seems · ▶ 6:18–6:24
We have the same 4500 watts that we did before, but now it’s up here
A simple water heater is more clever than it seems · ▶ 9:12–9:17

节省的能源到底在边界哪一侧?

遮阳篷先把太阳辐射拦在玻璃外,空调则在热进入以后把它搬出去。换街灯的例子提供另一项比较条件:除了照明能耗,不同光谱对天文观测的影响也不同。把改造前后要完成的工作先写在纸上,分别核对采光、舒适、能耗和外溢影响,才知道同一个节能百分比能否回答自己的问题。

原文与时间戳 · 4
the solar heat gain coefficient.
Awnings: a simple cooling tech we apparently forgot about · ▶ 9:57–10:03
a lower SHGC dropped cooling demand a pretty similar amount to adding overhangs.
Awnings: a simple cooling tech we apparently forgot about · ▶ 10:37–10:47
55% of the energy of the high pressure sodium lamps it replaced.
The LED's Challenge to High Pressure Sodium · ▶ 3:22–3:30
their nearly monochromatic light output can easily be filtered out
The LED's Challenge to High Pressure Sodium · ▶ 10:56–11:01

机器省下的功耗,是否转成了使用者的工作?

卧式冷柜减少开门时的冷空气流失,但需要手动除霜的型号另有停机成本。算一次节能改造要先定下比较窗口:只数插座上一小时的功率,与把结霜、清空、除霜和重新降温的一整个周期算进去,比较的对象并不相同。

原文与时间戳 · 2
this is a manual defrost freezer
Chest Freezers; What they tell us about designing for X · ▶ 10:30–10:36
it must have been replaced by the warmer, less dense air in the room.
Chest Freezers; What they tell us about designing for X · ▶ 11:45–11:52

把一小时移到傍晚,真正改变了哪些负荷?

照明用电减少时,空调负荷仍可能向另一个方向变化;家庭设备和生活作息也会随年代改变。调时后一个月和调时前一个月的电费同时受天气、季节与技术影响。另一项周三休息的提议让人看清同一种比较错误:先数总工作时间、再数连续工作天数、最后数能和同事协作的时段。把不同结果揉成一个「时间安排更好」会丢掉最可测试的部分。

原文与时间戳 · 4
Another problem when trying to study daylight saving time
You Lose an Hour of Sleep to Save $4 · ▶ 2:31–2:35
rapid changes in technology and electrical use.
You Lose an Hour of Sleep to Save $4 · ▶ 2:31–2:36
Weekend Wednesday!
Two Is the Wrong Number of Days Off · ▶ 0:49–0:51
Weekend Wednesday won't work for the whole world
Two Is the Wrong Number of Days Off · ▶ 1:46–1:49

账本有盈余,改变制度就能省下同样一笔钱吗?

王室财政视频先取某一时期的地产收入和支出求差,随后提出另一种制度安排下能保留多少收入的问题。第一个数字是既有安排的结果,第二个问题需要说明谁在替代安排下控制资产、收取收益、负担维护。把同一项收入完整计入两种互斥情形,是这类比较中很容易犯的错误。

原文与时间戳 · 2
let’s compare their profits
The True Cost of the Royal Family Explained · ▶ 1:03–1:07
keep 100% of the revenue.
The True Cost of the Royal Family Explained · ▶ 1:47–1:51

结果出来前,什么会让你改主意?

Muller 用预先评审研究假设和方法,并在研究遵守方案的前提下承诺发表,无论结果如何。这个做法能限制看过数据以后只报告显著结果。Olson 则回顾一次股东表决。他说,社群原本期待票数提供证据,后来把讨论转向为什么这种证据没有出现。把这两段放在一起,可以检查两件事:结果出来前,有没有说过什么结果会削弱主张?解释改变后,还能提出哪项可以检验的新预测?

原文与时间戳 · 2
submitting hypotheses and methods for peer review before conducting experiments
Is Most Published Research Wrong? · ▶ 10:37–10:43
Whatever explains away the negative test result.
This is Financial Advice · ▶ 77:42–77:45

读完再想一步

拿四张概率不同的卡片,为它们设计两个是非问题。先比较两问可能得到的信息,再假设答案是「是」重新计算。两次计算用了同一个分母吗?