AI 与机器学习2018drive.google.com
Jeff Dean 的本科毕业论文：神经网络（1990）
Google’s Jeff Dean’s undergrad senior thesis on neural networks (1990)
进阶 · 约 40 分钟读完前置知识：神经网络基础、线性代数、梯度下降
导读
这篇论文是 Jeff Dean 1990 年的本科高年级论文，围绕神经网络训练展开；从评论看，其突出特点是问题意识超前、表达异常简洁清晰。具体方法与结论需读原文确认。
HN 关注点：Jeff Dean 的早年工作，也借机讨论学术写作是否容得下简洁表达，以及他后来在 TensorFlow 和编译器优化上的思路延续。
HN 评论精选
这几条评论里：支持 1 补充 3
支持赞叹 Jeff Dean 早年就能简洁高效地解决复杂问题。
一如既往，Jeff Dean 令人肃然起敬。他早年就处理了一个至今仍相关的问题，取得好结果，并清晰简洁地描述解决方案。我的硕士论文约 60 页，有用程度可能只有这篇的千分之一。
halflings查看英文原文
As always, Jeff Dean doesn't fail to inspire respect. Tackling a complex problem (still relevant today) at an early age, getting great results and describing the solution clearly/concisely. My master thesis was ~60 pages long, and was probably about 1/1000 as useful as this one.
补充讨论学术写作生态对简洁表达的容忍度，认为 Jeff Dean 也受益于院系环境。
一个被低估的方面是：找到一个愿意让你把本科论文写得这么简洁的院系。我的经验（主要是在研究生阶段）是，任何给我改稿的人都想要更多字数；哪怕一句话就能说清，也不被接受，必须写更多。Jeff Dean 是极出色的沟通者，但他也受益于被允许、甚至被鼓励优先考虑有效而简洁的表达。多数人没这么幸运，反而会学到这种简洁不讨好，别人会以为你在装懂或没做足文献功课。
mlthoughts2018查看英文原文
An underappreciated aspect of this is finding an academic department that would allow you to submit something this concise as a senior thesis. My experience, mostly in grad school, was that anyone editing my work wanted more verbiage. If you only needed a short, one-sentence paragraph to say something, it just wasn’t accepted. There had to be more. Jeff Dean is an uncommonly good communicator. But he also benefited from being allowed, perhaps even encouraged, to prioritize effective and concise communication. Most people aren’t so lucky, and end up learning that this type of concision will not go over well. People presume you’re writing like a know-it-all, or that you didn’t do due diligence on prior work.
补充指出论文主题与 TensorFlow 的图复制/编译器优化思路的延续，并补充个人论文托管经历。
Dean 的本科论文是关于训练神经网络的，而且主要选择涉及图内复制（in-graph replication），这并不完全令人意外；这仍是今天 TensorFlow 的一大问题之一。多数人没意识到的是，Dean 本质上是擅长编译器优化的计算机科学家，而 TF 基本上是把神经网络加速转化为编译器优化问题的尝试。我想感谢本科母校把我的本科论文托管了 25 年，只改过一两次 URL。有趣的是 Latex2Html 基本撑了 25 年和几次 URL 变更。底层主题（训练二分类器权重系数以最大化性能）仍与我今天的工作相关，尽管我当时还不懂梯度下降和 softmax。
dekhn查看英文原文
I guess it's not totally surprising that Dean's undergrad thesis was on training neural networks and the main choice was between or in-graph replication. This is still one of the big issues with TensorFlow today. One thing most people don't get is that Dean is basically a computer scientist with expertise in compiler optimizations, and TF is basically an attempt at turning neural network speedups into problems related to compiler optimization. I'd like to thank my undergrad university for hosting my undergrad thesis for 25 years with only 1-2 URL changes. Some interesting details include: Latex2Html held up, mostly, for 25 years and several URL changes. The underlying topic is still relevant (training the weight coefficients of a binary classifier to maximize performance) to my work today, even if I didn't understand gradient descent or softmax at the time.
补充好奇论文是否有导师指导，还是 Jeff Dean 独立完成。
想知道当时他的导师是谁，因为论文里似乎没提。或者他是自己独立完成的——不过话说回来，这也不意外。
mi_lk查看英文原文
Wonder who was his advisor back then, because I don't think it's mentioned in the thesis. Or he did this on his own, which is not surprising by the way.
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
- 2018Hacker News 讨论帖413 分 · 98 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