AI 与机器学习2018ftp.cs.ucla.edu编辑精选
因果推理的七根支柱及对机器学习的反思
Seven Pillars of Causal Reasoning with Reflections on Machine Learning
进阶 · 约 90 分钟读完前置知识：概率图模型基础、因果推断基本概念（do 算子/反事实）、机器学习基础
导读
Judea Pearl（图灵奖得主、因果推断奠基人）总结因果推理的七大支柱，并借此反思当前机器学习的局限：他认为现有 ML 缺乏因果模型与反事实推理能力，难以实现真正的智能。论文系统阐述了因果层级、干预、反事实等核心概念在科学推理中的作用。
HN 关注点：Pearl 对深度学习'只做曲线拟合'的批评引发 HN 读者热议，讨论焦点是因果推理与数据驱动的统计学习能否结合、以及因果推理是否真的是通向人类级智能的必要条件。
后来怎么样
Pearl 的因果推断框架（do-演算、结构因果模型）已成为因果发现与因果机器学习研究的标准语言，2018 年后催生了大量因果表征学习与因果 ML 的后续工作；本书/论文也是该领域被广泛引用的纲领性文献。
HN 评论精选
这几条评论里：质疑 2 补充 2
质疑质疑因果推理是否为人类智能的原因而非结果，并以语音识别为例说明人工归纳的过度正则化可能有害
Pearl 的论文是对逻辑推理与智能的扎实概括，但要警惕：连人类自己也不擅长泛化和正则化。值得思考的是，逻辑推理究竟是导致人类智能的原因，还是复杂统计现象的结果？一个人类级 AI 系统或许会包含全部七大支柱，但这不意味着这些推理能力是被专门编程进去的。类似的过度正则化例子可见于语音技术：语言学家把英语归纳为 50~60 个音素，但真实语音中的音素要多几个数量级，强迫识别系统按 60 个音素分类会丢失重要细节。这正是 WaveNet 等端到端系统表现更好的原因——抛弃过度正则化的概念后，模型可以更自由地学习。
yonkshi查看英文原文
Pearl's paper is a solid generalization on logical reasoning and intelligence, but we have to be wary that even us, humans, aren't perfect at generalization and regularization. Without going too meta, we have to contemplate whether logical reasoning causes human level intelligence or it's a mere effect of a complex statistical phenomenon. It may be the case that a human-level AI system will contain all seven pillars of causal reason, but it does not mean that these reasoning were specifically programmed into the AI. A similar example of this human over-regularization can be found speech technology, linguists have generalized 50~60 phonemes in English speech, though in real life speech, there are magnitudes more phonemes in play, if we force our speech recognition systems to classify speech into 60 phonemes, we miss out on important details. This is why end to end systems such as WaveNet is so much better than older systems, because when we ditch our over-regularized concept, it can learn more freely.
补充指出论文遗漏了强化学习，而 RL（如 Dyna）正是因果推理与统计学习结合最自然的领域
我很惊讶 Pearl 在文中完全没提强化学习。这个领域正是 Pearl 式因果推理框架与数据驱动的统计机器学习方法最接近交汇之处。Dyna 这类方法可以看作带有一定'推理'未来行动后果能力的强化学习。我怀疑在把因果方法嵌入规划方面还有很多工作要做，不过显然在 Atari 等领域已有不少进展。
howlin查看英文原文
I'm surprised Pearl doesn't mention reinforcement learning once in this essay. This domain is where Pearl-style causal reasoning frameworks most closely meets data-driven statistical machine learning methods. Methods such have Dyna (link below) can be seen as reinforcement learning with some degree of "reasoning" about future consequences of actions. I suspect that there is more work to be done in embedding causal methods into planning, though it's pretty clear that a lot of progress is being made in domains such as Atari. Dyna: https://www.cs.cmu.edu/afs/cs/project/jair/pub/volume4/kaelbling96a-html/node29.html
补充提出程序归纳或许能自动学习因果图，绕过 Pearl 方法需预先定义因果图的限制
我的直觉是'程序归纳'（program induction）方法有可能让因果学习成为可能——而不必像 Pearl 的技术那样预先定义因果图。这里是我一次失败尝试的描述（链接）。
BenoitEssiambre查看英文原文
My intuition is that a "program induction" approach has the potential make causal learning possible (without having to predefine the causal graph as I believe Pearl requires in his techniques). Here is a description of my failed attempt: https://www.quora.com/What-deep-learning-ideas-have-you-tried-that-didnt-work/answer/Benoit-Essiambre
质疑质疑科学发现的推理过程（开普勒/牛顿/爱因斯坦）能否被形式化为因果推理
开普勒、牛顿、爱因斯坦是靠什么推理得出引力定律的？是因果推理还是逻辑推理？开普勒的推理和牛顿、爱因斯坦的属于同一'范畴'吗？不确定这个问题能否被形式化——哥德尔定理似乎表明它不能。
resource0x查看英文原文
What kind of reasoning allowed Kepler/Newton/Einstein to derive the laws of gravitation? Was it causal reasoning? Logical reasoning? Does Kepler's reasoning belong to the same "category" as Newton's? Einstein's? Not sure this question can even be formalized. Godel theorem seems to show it can't.
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。
- 视频与演讲
历次讨论
- 2018Hacker News 讨论帖158 分 · 12 评
同领域还值得读
DeepSeek-R1：通过强化学习激励 LLM 的推理能力
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL
研究如何用强化学习激励大语言模型的推理能力，提出 DeepSeek-R1，并探讨 RL 训练对模型推理表现的影响。
AI 与机器学习2025 · 1,351 分 · 1,056 评 · 引用 30进阶 · 约 80 分钟编辑精选1-bit LLM 时代：面向低成本计算的三值参数
The Era of 1-bit LLMs: ternary parameters for cost-effective computing
提出用三值参数 {-1, 0, 1} 替代 LLM 中的浮点权重，即所谓“1-bit LLM”。
AI 与机器学习2024 · 1,040 分 · 447 评 · 引用 19进阶 · 约 45 分钟编辑精选DeepSeek-V3.2：推进开放大语言模型的前沿
DeepSeek-v3.2: Pushing the frontier of open large language models
这是 DeepSeek-V3.2 的技术报告，延续了 DeepSeek 开源权重模型的路线，目标是让开放模型继续逼近甚至追平前沿闭源模型。
AI 与机器学习2025 · 982 分 · 465 评进阶 · 约 90 分钟编辑精选通过自我对弈与通用强化学习掌握国际象棋和将棋（AlphaZero）
Mastering Chess and Shogi by Self-Play with General Reinforcement Learning
DeepMind 的 AlphaZero 论文：用通用强化学习算法（自我博弈 + MCTS + 神经网络），从零开始仅数小时训练便在国际象棋和将棋上击败了世界顶级引擎 Stock…
AI 与机器学习2017 · 539 分 · 270 评 · 引用 13进阶 · 约 90 分钟编辑精选用强化学习发现更快的矩阵乘法算法
Discovering faster matrix multiplication algorithms with reinforcement learning
论文提出 AlphaTensor，用强化学习搜索矩阵乘法算法。
AI 与机器学习2022 · 518 分 · 111 评专业 · 约 90 分钟编辑精选