AI 与机器学习2018arxiv.org
深度学习：一份批判性评估
Deep learning: a critical appraisal
入门 · 约 60 分钟读完前置知识：深度学习基础、机器学习基本概念
导读
Gary Marcus 在这篇论文中对深度学习进行了批判性评估，指出其若干局限：数据饥渴、迁移能力有限、难以整合先验知识等，并认为要达到通用人工智能（AGI），深度学习必须与其他技术结合。文章引发了对深度学习是否被过度炒作的讨论。
HN 关注点：对深度学习热潮的反思与质疑，讨论焦点集中在 AGI 目标是否合理、论文指出的局限是否已被 meta-learning 等新研究解决。
后来怎么样
这篇论文在 2018 年引发了对深度学习局限性和炒作问题的广泛讨论，Marcus 关于混合 AI 系统的观点成为后续 AI 路线之争（端到端学习 vs 神经符号结合）的代表性立场，常被双方引用。
HN 评论精选
这几条评论里：质疑 3 补充 1
质疑论文指出的局限已被 meta-learning 等新研究方向部分解决，作者忽略了这些进展
论文中的许多担忧，尤其是深度学习数据饥渴、迁移能力有限和整合先验知识的问题，已经被 meta-learning 和 learning-to-learn 领域的最新论文解决（有监督和强化学习场景都有）。元强化学习方面的近期研究甚至表明这种机制与大脑的学习方式非常相似。我们小组最近发表了一篇关于脉冲神经网络中 learning-to-learn 的论文，网络能学到任务族的先验，并在元强化学习中自动学到有用的探索策略。我不敢说这是通往 AGI 的正确道路，但这是深度学习研究中一个非常有前景的新方向，而这篇论文似乎忽略了它。
trextrex查看英文原文
Many concerns in this paper, especially about deep learning being data hungry, limited capacity for transfer, and integrating prior knowledge are addressed by recent papers in meta-learning and learning-to-learn [1,2,3 and many many others] both in the supervised and reinforcement learning contexts. In the case of meta-reinforcement learning there has been recent work [5] which seems to indicate that this mechanism is very similar to how learning works in the brain. In fact, my group recently published a paper [4] about learning-to-learn in spiking networks (making it biologically more realistic) and showing that the network learns priors on families of tasks for supervised learning, and learns useful exploration strategies automatically when doing meta-reinforcement learning. While I don't claim this is the right path to AGI, it's a very promising and new direction in deep learning research which this paper seems to ignore. [1] http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.5.323 [2] https://arxiv.org/abs/1611.05763 [3] https://arxiv.org/abs/1703.03400 [4] https://arxiv.org/abs/1803.09574 [5] https://www.nature.com/articles/s41593-018-0147-8 (Preprint: https://www.biorxiv.org/content/early/2018/04/06/295964)
质疑质疑追求 AGI 的意义，认为对社会而言这一目标并不必要
为什么目标是 AGI？从攀登珠峰的角度我能理解，但对整个社会而言呢？我们已经有 70 多亿个通用智能了。目标是要取代人类吗？作为人类，我不赞成这样做。
crazybit查看英文原文
Why not praise DL and strive for practical applications, both present and future, instead of having the goal and evaluating success based on some nebulous "artificial general intelligence"?
补充AGI 概念误导性地将心智简化为算法，忽视哲学层面对知识与智能的理解
我觉得 AGI 这个词很有误导性。它暗示人类心智的奇妙之处——知识、情感、感觉等——可以被量化并变成算法。很多 AI 研究者应该补一补科学哲学和不同知识形式的课——episteme、phronesis、techne——或许还有科学革命的结构（托马斯·库恩）。整个 AI 范式预设了通用智能可以脱离生物、情感、身体/感官而存在。这是 70 年代 AI 寒冬的原因之一，当时研究者把语言和人类知识简化为符号的算法操作。推荐从约翰·塞尔的'中文房间'论证入门。
hestefisk查看英文原文
The term AGI I find quite misleading. It promotes the idea that the wonders of the human mind, knowledge, feelings, sensations etc can be quantified and turned into something algorithmic. I think a lot of AI researchers could do with a basic introduction to philosophy of science and the different forms of knowledge —- episteme, phronesis, techne — and perhaps also the structure of scientific revolutions (Thomas Kuhn, notably). The whole paradigm of AI presupposed that generalised intelligence can exist without biology, feelings, a body / senses. This is one of the reasons that AI reached the so-called “AI winter” in the 70ies where researchers boiled language and human knowledge down to algorithmic manipulation of symbols. EDIT: John Searle’s Chinese Room argument is a good (and fun) place to start —- https://en.wikipedia.org/wiki/Chinese_room
质疑论文的担忧虽合理，但都是活跃研究课题且已有基于深度学习的解决路径
论文中几乎所有担忧都是活跃的研究课题，并且已有使用某种深度学习方法的解决方案。没人否认这些担忧是合理的，但也没人会说当前深度学习研究获得的知识对未来无用。比如关于数据饥渴：可以说动物/人类在低层面也接收大量视听输入，进化本身用数百万年的数据做了'预训练'；此外还有无监督学习、强化学习，以及 one-shot/zero-shot/few-shot learning 和 meta-learning 等活跃研究方向。关于迁移能力有限：transfer-learning、meta-learning 和 multi-task-learning 正是处理这个问题的活跃领域。
albertzeyer查看英文原文
Almost all concerns in the paper are active research topics and do have certain solutions which do use some sort of deep learning approach. Depending on the viewpoint and interpretation, you could say that some of these approaches are hybrid solutions, but this is really just a matter of interpretation. No-one is really denying that the stated concerns are valid concerns. But also, no-one would say that the current knowledge gained from deep learning research will not be useful in the future. Of course, maybe for some aspects, you would need more radical new ideas, but I doubt that for future methods, nothing from the current methods will be used in some way. E.g.: 3.1. Deep learning thus far is data hungry. First, you could argue that on a low-level, an animal/human gets quite a lot of visual and audio input, so it's data hungry as well. Then, you could argue that the evolution did already do some sort of pretraining/pre-wiring which helps, using million of years of data. Then, related to this is the topic of unsupervised learning and reinforcement learning. Then, dealing with the aspect of learning with small amounts of data, there are the active research topics of one-shot-learning, zero-shot-learning of few-shot-learning. Related is also meta-learning. 3.2. Deep learning thus far is shallow and has limited capacity for transfer. Transfer-learning, meta-learning and multi-task-learning are active research areas which deal with this. 3.3. Deep learning thus far has no natural way to deal with hierarchical structure. There are various approaches also for this. This is also an active research area. 3.4. Deep learning thus far has struggled with open-ended inference. This is also an active research area. 3.5. Deep learning thus far is not sufficiently transparent. Also this is an active research area. And then, you could also argue that the biological brain also suffers at this. 3.6. Deep learning thus far has not been well integrated with prior knowledge. This is also an active research area. Etc.
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85 分 · 18 评
- 2018Hacker News 讨论帖147 分 · 61 评
- 2019Hacker News 讨论帖77 分 · 41 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