故事 · 一篇死后才发表的遗稿
Story · A Manuscript Published After His Death
18 世纪英国长老会牧师 Thomas Bayes 一生研究概率,却从未发表这条以他命名的定理。1761 年他去世后,
好友 Richard Price 整理遗稿,于 1763 年代为发表。两个多世纪后,这条"用新证据更新旧信念"的公式,
竟成了现代世界的隐形引擎 —— 医学诊断、垃圾邮件过滤、自动驾驶、大语言模型的推断,背后都是 Bayes 当年的思想:
后验 = 先验 × 似然,再归一化。
Thomas Bayes, an 18th-century English Presbyterian minister, spent his life studying probability — and never published the theorem now bearing his name.
After his death in 1761, his friend Richard Price dug through the manuscripts and published the paper for him in 1763.
More than two centuries later, that simple rule for "updating old beliefs with new evidence" turned into the hidden engine of the modern world —
medical diagnosis, spam filtering, self-driving cars, large language model inference, all run on what Bayes wrote down:
posterior = prior × likelihood, then renormalize.
1 把 10000 人摊在面前数一数 Lay 10 000 People in Front of You and Count
自然频率natural frequency每个小方块代表 1 个人。深红=真患病且阳性,橙=健康却误报阳性,浅灰=正确判阴。阳性人群 = 深红 + 橙,里面深红的占比就是后验概率。 Each square is one person. Dark red = truly sick and positive, orange = healthy but falsely positive, pale gray = correctly negative. The whole positive crowd = red + orange — and red's share within that crowd is the posterior probability.
2 只看"阳性"这群人:真假各占几成 Zoom in on the Positives: Who's Real, Who's a False Alarm
把所有阳性的人拎出来排成一条 —— 真患者在左,虚惊者在右。病越罕见,橙色越长 —— 这就是为什么单看"准确率 99%"会骗人。基础率(先验)不可忽视。 Pull every positive patient out and line them up — true cases on the left, false alarms on the right. The rarer the disease, the longer the orange — that's why "99% accurate" can be deeply misleading on its own. Never ignore the base rate (the prior).
3 现实里的贝叶斯 Bayes in the Real World
医学检测:罕见病初筛阳性≠确诊,必须复检 —— 否则大量虚惊会带来过度治疗与恐慌。
Medical screening: a positive screen for a rare disease is not a diagnosis — confirmation testing is mandatory, otherwise the wave of false positives causes overtreatment and panic.
垃圾邮件过滤:贝叶斯过滤器根据词频不断更新"是垃圾"的后验概率,越用越准。
Spam filtering: Bayesian filters keep updating "this is spam" posteriors from word frequencies — the more mail you label, the sharper they get.
先验更新:每来一条新证据,今天的后验就是明天的先验,信念在迭代中逼近真相。
Bayesian updating: every new piece of evidence turns today's posterior into tomorrow's prior — beliefs converge toward truth through iteration.
基础率谬误:忽略事件本身有多罕见,是人脑最常犯的概率错误,贝叶斯把它扳正。
Base-rate neglect: ignoring how rare an event is in the first place is the human brain's most common probability mistake — Bayes' theorem snaps it straight.
一句话In One Line
人脑天生忽略基础率(先验):一听"99% 准"就把它当成患病概率,却忘了"病有多罕见"。
贝叶斯定理把它扳正 —— 后验 = 先验 × 似然,再归一化。罕见的事,即使证据强,也未必成真;常见的事,弱证据也足以确信。
把概率换成自然频率(10000 人里多少人),这个反直觉的悖论立刻一目了然 —— 这正是教医生读检测报告的标准方法。
The human brain instinctively ignores the base rate (the prior): hear "99% accurate" and you take it for the chance you're sick — forgetting how rare the disease actually is.
Bayes' theorem snaps it back: posterior = prior × likelihood, then renormalize. Rare events stay improbable even given strong evidence; common events are confirmed by weak evidence.
Recast the probabilities as natural frequencies ("X people out of 10 000") and the paradox becomes blindingly obvious — exactly the technique medical educators use to train clinicians to read test reports.
常见误用Common Mistakes
把准确率当成患病概率。P(阳│病) ≠ P(病│阳),两者会差很远(条件颠倒谬误)。
Treating test accuracy as the probability of being sick. P(positive | sick) ≠ P(sick | positive) — they can be wildly different (the inverse-conditional fallacy).
忽略先验(基础率)。病越罕见,阳性里假阳越多,先验必须纳入计算。
Ignoring the prior (base rate). The rarer the disease, the more of the positive crowd is false alarms — the prior must enter the calculation.
用模糊概率讲,听者算不清。改用自然频率"10000 人里…",悖论瞬间清晰。
Explaining with abstract probabilities — listeners can't do the math. Switch to natural frequencies ("out of 10 000 people…") and the paradox becomes instantly clear.