按年被提到
20142015 · 8 次2026
机器学习与深度学习›神经科学启发的 AI·arXiv:1410.5401 · cs.NE

Neural Turing Machines

神经图灵机
Alex Graves, Greg Wayne, Ivo Danihelka · 2014

给神经网络配可微外部记忆,学会简单算法

开始阅读 ↗PDFalphaXiv 讨论
被推荐 · 30 人首次 2014-10-21 · 最近 2026-10-30
30不同的人
25活跃月份
32讨论串
1,058被引(OpenAlex)

把神经网络与可通过注意力机制读写的外部记忆耦合,类比图灵机或冯·诺依曼架构,但整体端到端可微,因此可用梯度下降训练。初步结果表明,神经图灵机能从输入输出样例中推断出复制、排序和联想回忆等简单算法。

HN 为什么提它

评论者持续把它当作带外部记忆的神经网络和让网络学习算法的代表,用来讨论神经网络能否做算法推理,也承认其实验多为玩具问题。

奠基之作难度 4 / 5约 50 分钟经得起时间
读之前最好知道
1循环神经网络与 LSTM
2注意力机制
3图灵机与冯·诺依曼架构
4反向传播与梯度下降

学习路径

读这篇之前,按顺序读书架上的 4 篇 · 参考文献来自 OpenAlex,顺序和要点由 Claude Code 排出

这条路径帮你建立 LSTM 序列建模、编码器-解码器和注意力机制的基础,神经图灵机的读写头寻址正是软注意力的延伸。

  1. 1
    Generating Sequences With Recurrent Neural Networks用循环神经网络生成序列 · 2013

    神经图灵机的控制器常用 LSTM,这篇给出 LSTM 序列建模的基础。

    读理解 LSTM 如何通过隐状态做序列预测,作为对比:这类网络的记忆只存在于隐状态中。

  2. 2

    编码器-解码器把序列压进定长向量,这一瓶颈正是外部记忆要解决的。

    读抓编码器-解码器的结构和定长向量的信息瓶颈。

  3. 3
    Sequence to Sequence Learning with Neural Networks用神经网络进行序列到序列学习 · 2014

    seq2seq 展示了 RNN 学映射序列,也暴露了长序列表示能力的限制。

    读读模型结构与对长序列效果的讨论,与神经图灵机的复制任务对照。

  4. 4
    Neural Machine Translation by Jointly Learning to Align and Translate通过联合学习对齐与翻译实现神经机器翻译 · 2014

    软对齐的注意力给出可微的内容寻址思路,神经图灵机的读写寻址与之相通。

    读读对齐模型:打分、softmax 权重、加权求和得到上下文向量。

  5. 5
    Neural Turing Machines神经图灵机 · 2014 · 就是这篇

    用可微的读写头把外部记忆接到控制器上,端到端学简单算法。

    读重点读内容寻址与位置寻址、读写操作,以及复制、排序、联想回忆等实验。

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
推荐current_call · 2016-05

“You might enjoy this paper if you haven't already read it.”

推荐它,其实验展示了 Neural Turing Machines 能做什么
Programming with a Differentiable Forth Interpreter原评论 ↗
推荐cs702 · 2016-12

“Take a look at these two papers:”

推荐这两篇论文,其中 Neural Turing Machines 把记忆建模为实数向量而非比特
Operational calculus on programming spaces and generalized tensor networks原评论 ↗
推荐typon · 2024-04

“check out the NTM paper: Neural Turing Machines”

推荐看 NTM 论文,它给神经网络加了可微分的外部记忆
Transformer as a general purpose computer原评论 ↗
批评apsec112 · 2018-06

“But the paper was widely advertised in the mainstream press as being successful, even though it was not tested on”

Deep learning: a critical appraisal原评论 ↗
批评rvz · 2026-10

“This concept was hyped up (because DeepMind) for a while at the time and in reality it amounted to nothing.”

批评这个概念当年因 DeepMind 被炒作,实际上一事无成
Show HN: A working Nintendo in ~1GB of Qwen weights原评论 ↗
引用nhaehnle · 2014-10

“For arXiv submissions, please submit the abstract link; in this case: Neural Turing Machines”

提醒提交 arXiv 论文应提交摘要页链接,并给出这篇的摘要页
Neural Turing Machines [pdf]原评论 ↗
看全部 20 条 →
引用a-priori · 2015-02

“This is the basis of an architecture called a 'neural turing machine' and it is, indeed, Turing complete and can be trained through standard neural network training algorithms to perform arbitrary computations.”

拿它说明神经图灵机图灵完备,可用标准算法训练来执行任意计算
A Brief Overview of Deep Learning原评论 ↗
引用compbio · 2015-03

“A fairly recent addition to deep nets is making use of memory: Neural Turing Machines "Neural Turing Machines".”

举例说明深度网络开始利用外部记忆,可能缩小与人类的差距
Images that fool computer vision raise security concerns原评论 ↗
引用teraflop · 2015-05

“And we have separate evidence that they can (in principle) capture semantic meaning and algorithmic reasoning as well, for example: Neural Turing Machines”

引用它作为证据,说明神经网络原则上能做语义和算法推理
The Unreasonable Effectiveness of Recurrent Neural Networks原评论 ↗
引用deepnet · 2015-05

“based on Graves "Neural Turing Machines" Neural Turing Machines, which attempts to infer algorithms from the result.”

指出自己的后续工作基于 Graves 的这篇论文,尝试从结果中推断算法
The Unreasonable Effectiveness of Recurrent Neural Networks原评论 ↗
引用RoboTeddy · 2015-07

“For example, there's software that can learn some simple algorithms: Neural Turing Machines”

举它为例,说明已有软件能学会一些简单算法
The Cyborg Compulsion原评论 ↗
引用kastnerkyle · 2015-08

“Turing machines [8]”

在应用示例列表中引用 Neural Turing Machines
A Beginner’s Guide to Restricted Boltzmann Machines原评论 ↗
引用sxyuan · 2015-09

“One related work that came to mind was Neural Turing Machines, which augments the LSTM network with a (fixed size) memory [1].”

提到相关工作 Neural Turing Machines,给 LSTM 加了定长外部记忆
Inferring Algorithmic Patterns with Stack原评论 ↗
引用andreyk · 2016-01

“There is also research lately into weirder types of neural nets such as 'Neural Turing Machines'.”

举它为例,说明近来有更怪的神经网络,如带外部记忆的 Neural Turing Machines
Turing Machines Are Recurrent Neural Networks (1996)原评论 ↗
引用alexbeloi · 2016-05

“Here's the original paper on Neural Turing machines: Neural Turing Machines”

给出 Neural Turing Machines 原始论文,涉及把序列预测外推到比训练更长
One-shot Learning with Memory-Augmented Neural Networks原评论 ↗
引用sapphireblue · 2016-07

“In fact there is already a lot of solid work just on this subject:”

指出已有不少扎实工作研究从示例学习算法,这篇是其中之一
How AI based programming could work原评论 ↗
引用unlikelymordant · 2017-07

“How would you sort integers? using neural turing machines: Neural Turing Machines”

回答如何给整数排序:用 neural turing machines 这类记忆增强网络
The Limitations of Deep Learning原评论 ↗
引用stev0lution · 2020-02

“Preliminary results demonstrate that Neural Turing Machines can infer simple algorithms such as copying, sorting, and associative recall from input and output examples.”

引用其初步结果:带外部记忆的网络能从输入输出示例学会复制、排序和联想回忆
Fast Differentiable Sorting and Ranking原评论 ↗
引用nmfisher · 2020-06

“Are you aware of the Neural Turing Machine”

问对方是否了解它:可微读写的外部记忆架构,但多是玩具问题
GPT-3 and Scaling Trends原评论 ↗
引用somethingAlex · 2021-07

“Interesting related work: Graves[0] trained a "neural Turing machine" to encode instructions for a basic sorting algorithm.”

提到相关工作:Graves 训练神经图灵机学会基本排序算法
Fluxsort: A stable adaptive partitioning comparison sort原评论 ↗