“Check out the recent paper "Building Machines That Learn and Think Like People":”
Building Machines That Learn and Think Like People
认知科学视角:真正类人的 AI 需要因果模型、直觉理论和组合性
文章回顾认知科学的进展,指出基于端到端深度神经网络的系统虽在物体识别、电子游戏和棋类上达到甚至超过人类,但与人类智能在关键方面不同。作者主张类人机器应:构建支持解释与理解的世界因果模型;以直觉物理和直觉心理学为学习的基础;利用组合性和学会学习(learning-to-learn)快速习得并泛化知识,并提出结合神经网络与结构化认知模型的具体挑战和路线。
HN 上的人把它当作「当前 AI 方法为什么不足、类人学习应该怎样」的易读综述来推荐,多用于支持因果模型与类人推理的论点。
学习路径
读这篇之前,按顺序读书架上的 7 篇 · 参考文献来自 OpenAlex,顺序和要点由 Claude Code 排出先看深度学习的整体进展,再看这篇文章用来对照的几类系统:图像描述、带注意力的生成、结构化生成模型、外部记忆和分层 / 内在动机 RL,最后读认知科学视角的批评。
- 1
- 2
目标论文用图像描述系统说明网络能产出流畅句子却缺乏真正理解。
读CNN 编码加 LSTM 解码的结构;留意它的失败案例。
- 3
带视觉注意力的描述模型,是文中讨论的一个代表。
读软 / 硬注意力的做法和可视化。
- 4
DRAW 是逐步生成图像的神经模型,文中拿来和人类的概念生成作对比。
读注意力窗口加循环生成的框架;不必推导变分下界。
- 5
Attend, Infer, Repeat 把场景理解做成结构化生成模型,靠近目标论文主张的组合性与因果模型。
读逐步推断场景中物体个数与属性的框架。
- 6
- 7
分层强化学习结合内在动机,对应文中关于目标、分层与规划的讨论。
读元控制器与子目标控制器的分层结构。
- 8Building Machines That Learn and Think Like People
目标论文:从认知科学角度提出类人 AI 需要的核心要素。
读因果模型、直觉物理与直觉心理、组合性与学会学习几个核心主张,再看对深度网络现状的评论。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“See here for an excellent overview of limitations in our current approaches to AI: Building Machines That Learn and Think Like People”
“I also posted a link above which is an easily readable exposition on just how current approaches fall short.”
“Hey guys, I found the non-crappy papers:”
“Josh Tenenbaum has pushed the idea of causal models and inference, and I find it compelling”
“tbh I haven't read the bottom two”
看全部 9 条 →
“This is an important part of how to build models that think like people[1].”
“I suspect that the current generation of driverless vehicles will be the same. [LINK] Building Machines That Learn and Think Like People”
“Lastly, Building Machines That Learn and Think Like People lays out this vision in a lot more detail.”
- Building machines that learn and think like people ↗tonybeltramelli · 2016-04-04 · 16 分
- Building Machines That Learn and Think Like People ↗eli_gottlieb · 2016-04-04 · 6 分
- Building Machines That Learn and Think Like People ↗figure8 · 2016-04-05 · 2 分
- Building Machines That Learn and Think Like People ↗ansible · 2016-07-30 · 1 分
- Building Machines That Learn and Think Like People ↗tma-1 · 2017-02-10 · 2 分
- Building Machines That Learn and Think Like People ↗rsyngh · 2018-01-28 · 2 分