日期:2026-06-23 性质:前期文献核查笔记,非最终裁决报告 用途:为总课题「归纳问题与“无免费午餐”——为什么机器学习能泛化」提供哲学史与认识论基础。 接口说明:de Finetti / Dutch book / Solomonoff / NFL / PAC / VC 等已在 2026-06-22 概率与贝叶斯主义基础大体检 与 2026-06-12 语言作为压缩 中覆盖;本笔记只作接口引用,不重复展开。
一、关键来源
来源 1:Hume(1748 / 1739)论因果与归纳
出处:David Hume, An Enquiry Concerning Human Understanding(1748)§4.2;A Treatise of Human Nature(1739)1.3.6。 二手入口:SEP The Problem of Induction, §1 “Hume’s Problem”. URL: https://plato.stanford.edu/entries/induction-problem/
- 中文摘要:Hume 把因果推断追溯到“恒常联结”的经验,并追问:从“过去观察到的规律”到“未来/未观察案例也如此”这一推断,能否被理性证明?他区分“观念间关系”与“事实与存在”,指出任何证明要么先验(无法成立,因为自然进程改变不自相矛盾),要么经验(必预设“未来像过去”,循环论证)。最终把归纳归于“习惯/想象”,而非理性。
- 原始关键句:
- “All reasonings may be divided into two kinds, namely, demonstrative reasoning, or that concerning relations of ideas, and moral reasoning, or that concerning matter of fact and existence.”(Enquiry §4.2.18)
- “it implies no contradiction that the course of nature may change, and that an object seemingly like those which we have experienced, may be attended with different or contrary effects.”(Enquiry §4.2.18)
- “When the mind, therefore, passes from the idea or impression of one object to the idea or belief of another, it is not determin’d by reason, but by certain principles, which associate together the ideas of these objects, and unite them in the imagination.”(Treatise 1.3.6.12)
- “This principle is ‘custom’ or ‘habit’… an operation of the soul, when we are so situated, as unavoidable as to feel the passion of love, when we receive benefits.”(Enquiry §5.1.8)
- 章节/段落定位:Enquiry §4.1–4.2 提出 Uniformity Principle(UP)与两难;§5.1–5.2 给出“习惯/想象”的心理学替代方案;Treatise 1.3.6 是论证更完整版本。
- 证据标签:[文献较稳](SEP 与 Hume 原文多版本互核一致)
来源 2:Popper(1934/1959)证伪主义与对归纳的拒绝
出处:Karl Popper, Logik der Forschung(1934)/ The Logic of Scientific Discovery(1959)。 二手入口:SEP Karl Popper, §3 “The Problem of Demarcation” 与 §6 “Probability, Knowledge and Verisimilitude”. URL: https://plato.stanford.edu/entries/popper/
- 中文摘要:Popper 接受 Hume 对归纳的批判,并进一步主张“科学中从未真正使用归纳”。他以“可证伪性”作为科学与非科学的划界标准:全称命题无法被经验证实,但可被单一反例逻辑证伪;科学通过提出大胆猜想并经受严峻检验而进步,检验结果只提供“corroboration(确证)”而非概率意义上的确认。后来他引入 Tarski 式“verisimilitude(逼真度)”以说明科学向真理逼近,但该形式定义后被 Miller 与 Tichý 证明对假理论失效。
- 原始关键句:
- “Popper is unusual amongst contemporary philosophers in that he accepts the validity of the Humean critique of induction, and indeed, goes beyond it in arguing that induction is never actually used in science.”(SEP §3)
- “it is logically impossible to verify a universal proposition by reference to experience (as Hume saw clearly), but a single genuine counter-instance falsifies the corresponding universal law.”(SEP §3)
- “Every genuine scientific theory then, in Popper’s view, is prohibitive, because the theories of natural science take the form of universal statements.”(SEP §3)
- “a theory that has withstood rigorous testing should be deemed to have received a high measure of corroboration. and may be retained provisionally as the best available theory until it is finally falsified and/or is superseded by a better theory.”(SEP §3)
- 关于逼真度缺陷:Miller 与 Tichý(1974)独立证明“the conditions specified by Popper in his accounts of both qualitative and quantitative verisimilitude for comparing the truth- and falsity-contents of theories can be satisfied only when the theories are true.”(SEP §6)
- 章节/段落定位:Logic of Scientific Discovery 第 I–IV 章提出划界与证伪;§82–§83 讨论 corroboration;逼真度主要在 Conjectures and Refutations(1963)ch. 10。
- 证据标签:[文献较稳]
来源 3:Reichenbach(1938/1949)频率解释与“直规则”的实用辩护
出处:Hans Reichenbach, Experience and Prediction(1938);The Theory of Probability(1949)。 二手入口:SEP Hans Reichenbach, §2.3 “The Theory of Probability” 与 §3.2 “Mature Views: Experience and Prediction (1938)”. URL: https://plato.stanford.edu/entries/reichenbach/
- 中文摘要:Reichenbach 把概率解释为极限相对频率,并在此基础上提出“直规则”(straight rule):把观察到的相对频率直接当作总体极限频率的估计。他承认 Hume 论证不可反驳,但给出一种“实用辩护”:如果事件序列存在极限频率,直规则在极限下必收敛到该值;若世界毫无规律,则任何方法都无法预测,因此“如果未来可言说,直规则能带我们到达”。然而 SEP 也指出:任何“相对频率 + 一个趋于 0 的修正项”的方法都会收敛,所以直规则并不唯一;且该辩护对有限样本/收敛速度无话可说。
- 原始关键句:
- “Primary or fundamental inductive inference consists of taking observed relative frequencies as probabilities, that is, as limiting relative frequencies. This procedure, referred to as the ‘straight rule’, implies that one should take the current empirical distribution to resemble the limiting distribution.”(SEP §3.2)
- “The justification of such taking, or ‘positing’ in Reichenbach’s terminology, is that if there is a limiting relative frequency to a sequence, this procedure will converge to it.”(SEP §3.2)
- “if it is possible to make statements about the future we shall find them by means of this method.”(Reichenbach 1949: 475;转引自 SEP 相关讨论与二手综述)
- “He further acknowledges that any procedure that estimates the probability to be the relative frequency, plus any quantity that keeps the estimate between 0 and 1 and that itself converges to 0, will also converge to the limiting frequency if such exists.”(SEP §3.2)
- 章节/段落定位:Experience and Prediction §§37–39 讨论归纳与“posit”;The Theory of Probability §30 讨论 normal sequence 与收敛。
- 证据标签:[文献较稳](SEP 对直规则与实用辩护的叙述学界通用);Reichenbach 1949: 475 的逐字原文 未直接核取,[仍不确定/弱]
来源 4:Carnap(1950/1952/1963)归纳逻辑与 λ-连续统
出处:Rudolf Carnap, Logical Foundations of Probability(1950/1962);The Continuum of Inductive Methods(1952)。 二手入口:SEP Rudolf Carnap, supplement “Inductive Logic”;SEP Interpretations of Probability, §3.2.1 “The logical interpretation”;Carnap 1952 原版前言 PDF. URLs: https://plato.stanford.edu/entries/carnap/inductive-logic.html;https://plato.stanford.edu/entries/probability-interpret/;https://www.phil.cmu.edu/projects/carnap/editorial/latex_pdf/1952-1.pdf
- 中文摘要:Carnap 把“概率1”(确认度)视为逻辑/语义关系,把 Hume 的“自然齐一性”替换为语言框架内的对称性(structure-description 上的均匀分布)。最简单的确认函数 c* 对每个 structure-description 赋等概率,从而实现“从经验学习”。1952 年他进一步提出 λ-连续统,用单个参数 λ 调节先验权重与证据权重之比;λ→∞ 时接近无差异原则,λ→0 时接近纯频率主义。但该方案依赖语言选择,λ 的取值无法由逻辑单独决定;且在无限个体语言中,普遍定律的确认度恒为 0(被 Popper 与 Earman 批评)。
- 原始关键句:
- “confirmation or probability1 is framework-relative for Carnap… logical truths… are to be assigned the maximal value of (absolute) confirmation 1 whatever the evidence.”(SEP Carnap/Inductive Logic)
- “It is not claimed that c* is necessarily the best explicatum possible… It will not be claimed that c* is a perfectly adequate explicatum for probability1, let alone that it is the only adequate one.”(Carnap 1950b: ix, 563;SEP 引用)
- 关于 λ-连续统:
\[
c_{\lambda}(\text{individual } s+1 \text{ is } P_j,\ s_j \text{ of the first } s \text{ individuals are } P_j) = \frac{s_j + \lambda/k}{s + \lambda}
\]
“The higher the value of λ, the less impact evidence has: induction from what is observed becomes progressively more swamped by a classical-style equal assignment to each of the k possibilities.”(SEP Interpretations of Probability §3.2.1) - Carnap 1952 前言:“Essentially, one parameter is sufficient for a complete characterization of each method.”(Carnap 1952 PDF preface)
- “a universally quantified sentence may receive a probability of 0… this is one of the problems with Carnap’s account.”(SEP Carnap/Inductive Logic)
- 章节/段落定位:Logical Foundations §§53(概率公理)、§§94–96(对称性/二项律)、附录 §110(c* 定义);Continuum of Inductive Methods 全文定义 λ-连续统。
- 证据标签:[文献较稳]
来源 5:Mayo & Spanos(2006)/ Mayo(2018)严重检验
出处:Deborah G. Mayo & Aris Spanos, “Severe Testing as a Basic Concept in a Neyman–Pearson Philosophy of Induction”, British Journal for the Philosophy of Science 57(2):323–357(2006);Deborah G. Mayo, Statistical Inference as Severe Testing(CUP 2018)。 一手入口:Mayo & Spanos 2006 PDF;Mayo 2018 前言 PDF. URLs: https://errorstatistics.com/wp-content/uploads/2013/12/2006mayo_spanos_severe_testing.pdf;https://assets.cambridge.org/97811076/64647/frontmatter/9781107664647_frontmatter.pdf
- 中文摘要:Mayo 与 Spanos 把 Neyman-Pearson 的误差概率重新解释为“检验的严格性/探测性”(severity/probativeness),而非行为主义地接受/拒绝规则。一个推断被支持,当且仅当该假设通过了严峻检验:若假设为假,检验有很大概率会发现其错误。Mayo 2018 把这一思路扩展为对统计推断战争的回应,强调概率在科学中的作用是“评估并控制方法发现与避免错误解释的能力”,而非给假设赋后验概率。
- 原始关键句:
- “We argue that the relevance of error probabilities is to ensure that only statistical hypotheses that have passed severe or probative tests are inferred from the data.”(Mayo & Spanos 2006 Abstract)
- “A statistical hypothesis H passes a severe test T with data x0 if, (S-1) x0 agrees with H, and (S-2) with very high probability, test T would have produced a result that accords less well with H than x0 does, if H were false.”(Mayo & Spanos 2006 §1.3)
- “In the severe testing view, probability arises in scientific contexts to assess and control how capable methods are at uncovering and avoiding erroneous interpretations of data. That’s what it means to view statistical inference as severe testing.”(Mayo 2018 Preface, xii)
- “The severe testing perspective substantiates, using modern statistics, the idea Karl Popper promoted, but never cashed out.”(Mayo 2018 Preface, xii)
- 章节/段落定位:Mayo & Spanos 2006 §1.2–1.3 提出 severity rationale;§4 讨论 SIA/SIR;Mayo 2018 Preface 给出核心口号,Excursion 2 §2.3 “Popper, Severity, and Methodological Probability” 处理与 Popper 的关系。
- 证据标签:[文献较稳]
来源 6:现代贝叶斯辩护——核心规范与“先验问题”
出处:SEP Bayesian Epistemology(主要作者 Michael G. Titelbaum)。 URL: https://plato.stanford.edu/entries/epistemology-bayesian/
- 中文摘要:贝叶斯认识论把“概率”解释为合理置信度,核心规范是 Probabilism(置信度符合概率演算)与 Conditionalization(新证据通过条件化更新)。该框架能解释 Eddington 观测等现象,但在“枚举归纳”情形中,Probabilism + Conditionalization alone 无法决定置信度应上升还是下降——这完全取决于先验分布。因此“先验问题”是贝叶斯方案的关键缺口。主观贝叶斯主义允许任意相干先验;客观贝叶斯主义试图用无差异原则等约束先验,但原则本身充满争议。
- 原始关键句:
- “According to the first norm, called Probabilism, one’s credences… ought to be probabilistic… the second norm, called the Principle of Conditionalization.”(SEP §1.2)
- “Probabilism and the Principle of Conditionalization, alone, are too weak to entitle us to say whether one’s credence ought to change inductively or counter-inductively in the above example.”(SEP §1.5)
- “So, besides the coherence norms (such as Probabilism), are there any other norms that govern one’s prior? This is known as the problem of the priors.”(SEP §1.5)
- 章节/段落定位:SEP §1.2 提出 Probabilism + Conditionalization;§1.5 讨论先验问题;§1.6–1.7 讨论 Dutch Book、accuracy-dominance 等辩护。
- 证据标签:[文献较稳]
来源 7:de Finetti 表示定理与可交换性(接口点)
出处:Bruno de Finetti, Theory of Probability(1931/1964/1974)。 二手入口:SEP The Problem of Induction, §3.3 “Bayesian solution” 与 SEP Interpretations of Probability §3.2.1. URL: https://plato.stanford.edu/entries/induction-problem/
- 中文摘要:de Finetti 的表示定理证明:若一个无限序列的联合分布在排列下不变(可交换),则它可写成某个参数分布上的混合,数据在该参数下表现得像独立同分布抽样。可交换性可被视为 Hume“未来像过去”的一种形式化。但它本身是一个对称性/经验假设,并非从概率演算单独推出;主观贝叶斯主义者把它当作数据假设,Carnap 式客观贝叶斯主义者把它当作理性要求。
- 原始关键句:
- “De Finetti proved a general representation theorem that if the joint probability distribution of an infinite sequence of random variables is assumed to be exchangeable, then it can be written as a mixture of distribution functions from each of which the data behave as if they are independent random draws.”(SEP Problem of Induction §3.3)
- “The assumption of exchangeability may be seen as a natural formalization of Hume’s assumption that the past resembles the future.”(SEP Problem of Induction §3.3)
- 证据标签:[文献较稳]
来源 8:形式学习理论 / NFL / Solomonoff(仅接口)
- Wolpert & Macready (1997) “No Free Lunch Theorems for Optimization”, IEEE Transactions on Evolutionary Computation 1(1):67–82. DOI: 10.1109/4235.585893.
- Solomonoff (1964) “A Formal Theory of Inductive Inference”, Information and Control 7:1–22, 224–254.
- 这些已在 2026-06-22 概率与贝叶斯主义基础大体检 与 2026-06-12 语言作为压缩 中处理,本笔记只标记接口。
二、主要结论(带证据标签)
- Hume 的两难仍然硬
对“未来/未观察案例会像过去”这一 Uniformity Principle,既无法给出先验证明(否定它不矛盾),也无法给出经验证明(必预设自身)。[文献较稳]
- Popper 的证伪主义接受 Hume 批判,但付出了代价
科学推理被替换为“猜想—证伪—确证”;确证(corroboration)只记录过去表现,不保证未来可靠性;逼真度(verisimilitude)形式定义对假理论有缺陷。[文献较稳]
- Reichenbach 的实用辩护收窄了归纳问题
直规则在“极限频率存在”的前提下收敛;但它不是唯一收敛规则,且不提供有限样本或收敛速度保证。[文献较稳]
- Carnap 的归纳逻辑把“自然齐一性”语言化
c* 与 λ-连续统将对称性/无差异原则放进语言框架,使归纳学习成为分析性要求;但语言选择、λ 取值、无限域中普遍定律概率为 0、Putnam 形式学习理论批判,使其无法成为唯一或终极辩护。[文献较稳]
- Mayo 的严重检验提供了一条非贝叶斯、非 Popper 式的频率主义认识论路径
用误差概率评估“假设若为假,检验能否发现”;不依赖先验分布,但把“归纳”替换为“方法论的严格探测”。[文献较稳]
- 贝叶斯辩护的核心张力在于“先验问题”
Probabilism + Conditionalization 是形式真理,但单独无法推出枚举归纳;必须先验/可交换性/模型假设等额外输入。de Finetti 定理把主观概率与频率桥接起来,但桥本身不是无代价的。[文献较稳]
- 与 ML 泛化的接口
上述哲学线索共同说明:任何从有限样本到样本外泛化的推断,都必须额外假设(对称性/可交换性、模型类、先验/归纳偏置、世界结构)。NFL 定理把这一直觉形式化:没有偏置就没有泛化。PAC/VC 与 Solomonoff 分别在特定模型类或计算可学习性条件下给出边界。[理论整合]
三、与母裁决的接口
母裁决雏形要求:
- “归纳问题哲学上仍硬”:Hume 两难、Popper 对归纳的放弃、Reichenbach 实用辩护的狭窄性、Carnap λ-连续统的任意性、Bayesian 先验问题,共同支持这一点。
- “NFL / PAC / VC / Solomonoff 形式定理真而解释窄”:哲学史为这些定理提供了认识论前史;它们只是把“额外假设不可避免”形式化,而不是解决归纳问题。
- “真正答案是形式约束 + 隐式偏置 + 世界结构三足鼎立”:Carnap 的语言/对称性、Bayesian 的先验/可交换性、Mayo 的方法论约束,分别对应“形式约束”;DNN 的架构/优化器/数据增强对应“隐式偏置”;真实世界分布的规律性对应“世界结构”。
四、诚实缺口与待核点
- Hume 一手引文:本笔记中的关键句来自 SEP 提取;原始 Enquiry §4.2 / Treatise 1.3.6 的纸质/电子版页码未逐字核对。
- Reichenbach 1949: 475 名句:“if it is possible to make statements about the future…” 来自二手综述,原文 PDF 未直接获取。
- Carnap 1952 PDF:已获取前言(https://www.phil.cmu.edu/projects/carnap/editorial/latex_pdf/1952-1.pdf),但正文 λ-连续统的详细推导仍依赖 SEP 转述;LFP §110 c* 原文未逐字核对。
- Popper 一手引文:依赖 SEP Karl Popper;Logic of Scientific Discovery 全文未重新抓取。
- Mayo 2018 全书:仅获取前言/目录 PDF;正文 Excursion 2 的详细论证未完整阅读。
- Goodman 的 grue 问题:本笔记未深入;它进一步威胁 Carnap 的语言依赖方案,可在后续阶段补核。
- Cox 定理 / Joyce 准确度论证 / Dutch Book 细节:已在 2026-06-22 概率与贝叶斯主义基础大体检 中处理,本处只接口。
五、后续问题
- 是否需要把 Hume / Popper / Reichenbach / Carnap / Mayo 的关键段落补做一手页码核对?
- 子课题 B(PAC / VC / 形式学习理论)与子课题 C(深度学习泛化机制)应如何与本笔记的哲学线索精确接口?
- Goodman grue 问题与“projectible predicate”的现代处理(如 Lewis / Quine / Norton)是否要单独补一段?
- Mayo 的 severe testing 与深度学习中的“压力测试 / 分布外泛化评估”是否存在可直接借用的概念映射?