Attention Is All You Need(读论文)

在这里插入图片描述

Abstract

The dominant sequence transduction models are based on complex recurrent or
convolutional neural networks that include an encoder and a decoder.
The best
performing models also connect the encoder and decoder through an attention
mechanism.
We propose a new simple network architecture, the Transformer,
based solely on attention mechanisms, dispensing with recurrence and convolutions
entirely
.
主流的序列转换模型是基于复杂的循环神经网络(RNN)卷积神经网络(CNN),它们包含一个编码器和一个解码器。表现最好的模型还通过注意力机制将编码器和解码器连接起来。我们提出了一种新的简单的网络架构——Transformer,它完全基于注意力机制,完全舍弃了循环和卷积。

主要讲主流的翻译模型都是RNN和CNN,作者提出了一个全新的网络架构,彻底舍弃了RNN和CNN,这个新的架构叫Transformers

Experiments on two machine translation tasks show these models to
be superior in quality while being more parallelizable and requiring significantly
less time to train. Our model achieves 28.4 BLEU on the WMT 2014 Englishto-German translation task, improving over the existing best results, including
ensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task,
our model establishes a new single-model state-of-the-art BLEU score of 41.8 after
training for 3.5 days on eight GPUs, a small fraction of the training costs of the
best models from the literature. We show that the Transformer generalizes well to
other tasks by applying it successfully to English constituency parsing both with
large and limited training data.

在两个机器翻译任务上的实验表明,这些模型在质量上更优,同时具有更好的并行性,且训练所需时间显著减少。我们的模型在WMT 2014英译德翻译任务上达到了28.4的BLEU分数,比现有的最佳结果(包括集成模型)提高了超过2个BLEU。在WMT 2014英译法翻译任务中,我们的模型在八块GPU上训练3.5天后,建立了新的单模型最先进BLEU分数41.8,只用了文献中最佳模型训练成本的一小部分。我们通过在大规模和有限训练数据上成功将其应用于英语成分句法解析,展示了Transformer对其他任务的良好泛化能力。

这里主要讲这个Transformer架构不仅效果好,还省运算量,泛化能力还强:)
  1. 序列转换模型(Sequence Transduction Models)
  2. 循环神经网络(Recurrent Neural Networks,RNN)
  3. 卷积神经网络(Convolutional Neural Networks,CNN)
  4. 注意力机制(Attention Mechanism)
  5. 网络架构(Network Architecture)
  6. 舍弃了循环和卷积(Dispensing with Recurrence and Convolutions)
  7. 并行性 / 并行化(Parallelizable)
  8. 训练成本(Training Costs)
  9. BLEU分数(BLEU Score)
  10. WMT 2014英译德翻译任务(WMT 2014 English-to-German Translation Task)
  11. 集成模型(Ensembles)
  12. 单模型最先进水平(Single-Model State-of-the-Art)
  13. 英语成分句法解析(English Constituency Parsing)
  14. 泛化能力(Generalizes Well)
  15. 大规模训练数据(Large Training Data)
  16. 有限训练数据(Limited Training Data)
  17. 机器翻译任务(Machine Translation Tasks)

Equal contribution

~~Equal contribution. Listing order is random. Jakob proposed replacing RNNs with self-attention and started the effort to evaluate this idea. Ashish, with Illia, designed and implemented the first Transformer models and has been crucially involved in every aspect of this work. Noam proposed scaled dot-product attention, multi-head attention, and the parameter-free position representation and became the other person involved in nearly every detail. Niki designed, implemented, tuned, and evaluated countless model variants in our original codebase and tensor2tensor. Llion also experimented with novel model variants, was responsible for our initial codebase, and efficient inference and visualizations. Lukasz and Aidan spent countless long days designing various parts of and implementing tensor2tensor, replacing our earlier codebase, greatly improving results and massively accelerating our research.
同等贡献。列出的顺序是随机的。Jakob 提出了用自注意力替代 RNN 的想法,并开始评估这一想法的工作。Ashish 与 Illia 一起设计并实现了第一个 Transformer 模型,并在这项工作的各个方面都发挥了关键作用。Noam 提出了缩放点积注意力、多头注意力和无参数的位置表示,并成为参与几乎每一个细节的另一位主要贡献者。Niki 在我们最初的代码库和 tensor2tensor 中设计、实现、调试和评估了无数的模型变体。Llion 也尝试了新的模型变体,负责我们的初始代码库,以及高效的推理和可视化。Lukasz 和 Aidan 花费了无数个长日子设计并实现了 tensor2tensor 的各个部分,取代了我们早期的代码库,大大改善了结果并极大地加速了我们的研究。

这里讲几位作者的分工,可以略过

Work performed while at Google Brain.
Work performed while at Google Research.
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.~~

† 这项工作是在 Google Brain 期间完成的。
‡ 这项工作是在 Google Research 期间完成的。
第31届神经信息处理系统会议(NIPS 2017),美国加利福尼亚州长滩。
疑难词汇或短语:

  1. Equal contribution(同等贡献)
  2. Self-attention(自注意力)
  3. Scaled dot-product attention(缩放点积注意力)
  4. Multi-head attention(多头注意力)
  5. Parameter-free position representation(无参数的位置表示)
  6. Tensor2tensor(Tensor2Tensor,谷歌开源的机器学习框架,论文代码实现有github,可惜是tensorflow架构)
  7. Google Brain (谷歌大脑部门)
  8. Google Research (谷歌研究院)
  9. 31st Conference on Neural Information Processing Systems (NIPS 2017)(第31届神经信息处理系统会议)
  10. Long Beach, CA, USA(美国加利福尼亚州长滩市)

1 Introduction

Recurrent neural networks, long short-term memory [13] and gated recurrent [7] neural networks
in particular,
have been firmly established as state of the art approaches in sequence modeling and
transduction problems such as language modeling and machine translation [35, 2, 5]. Numerous
efforts have since continued to push the boundaries of recurrent language models and encoder-decoder
architectures [38, 24, 15].

循环神经网络,特别是长短期记忆网络(LSTM)[13] 和门控循环神经网络(GRU)[7],已经被确立为序列建模和转换问题(如语言建模和机器翻译)[35, 2, 5]的最先进方法。自那以来,众多努力持续推动着循环语言模型和编码器-解码器架构的边界 [38, 24, 15]。

先说RNN目前最牛的,LSTM和GRU是RNN的杰出代表

Recurrent models typically factor computation along the symbol positions of the input and output
sequences. Aligning the positions to steps in computation time, they generate a sequence of hidden
states ht, as a function of the previous hidden state ht−1 and the input for position t. This inherently
sequential nature
precludes parallelization within training examples, which becomes critical at longer
sequence lengths, as memory constraints limit batching across examples
. Recent work has achieved
significant improvements in computational efficiency through factorization tricks [21] and conditional
computation [32], while also improving model performance in case of the latter.
The fundamental
constraint of sequential computation, however, remains.
循环模型通常沿着输入和输出序列的符号位置分解计算。将位置与计算时间的步骤对齐,它们生成一系列隐藏状态 ( h_t ),作为前一隐藏状态 ( h_{t-1} ) 和位置 ( t ) 的输入的函数。这种本质上的序列特性排除了训练样本内的并行化,在更长的序列长度下,这变得至关重要,因为内存限制了跨样本的批处理。最近的工作通过分解技巧 [21] 和条件计算 [32] 在计算效率上取得了显著的改进,同时在后者的情况下还提高了模型性能。然而,序列计算的基本限制仍然存在。

但是!无法并行训练。虽然有人尝试解决RNN的并行化,但是还是有原理上的限制

Attention mechanisms have become an integral part of compelling sequence modeling and transduction models in various tasks, allowing modeling of dependencies without regard to their distance in
the input or output sequences [2, 19]. In all but a few cases [27], however, such attention mechanisms
are used in conjunction with a recurrent network.

注意力机制已经成为各种任务中引人注目的序列建模和转换模型的不可或缺部分,允许在输入或输出序列中不考虑距离地建模依赖关系 [2, 19]。然而,除少数情况外 [27],这种注意力机制都是与循环网络结合使用的。

attention早就应用在翻译任务中,不是作者们创新的,但是都是和RNN一起用

In this work we propose the Transformer, a model architecture eschewing recurrence and instead
relying entirely on an attention mechanism to draw global dependencies between input and output.

The Transformer allows for significantly more parallelizationand can reach a new state of the art in
translation quality
after being trained for as little as twelve hours on eight P100 GPUs.
在这项工作中,我们提出了 Transformer,一种避开循环、完全依赖注意力机制来在输入和输出之间建立全局依赖关系的模型架构。Transformer 允许显著更多的并行化,并且仅在八块 P100 GPU 上训练 12 小时后,就能在翻译质量上达到新的最先进水平。

再次说新的架构transformer,并且完全依赖注意力机制,抛弃RNN。transformer效果好,训练速度快,不依赖序列长度可以并行化
  1. Recurrent neural networks(循环神经网络)
  2. Long short-term memory (LSTM)(长短期记忆网络,RNN的一种)
  3. Gated recurrent neural networks (GRU)(门控循环神经网络,RNN的一种)
  4. Sequence modeling(序列建模)
  5. Transduction problems(转换问题)
  6. Factor computation along the symbol positions(沿符号位置分解计算)
  7. Aligning the positions to steps in computation time(将位置对齐到计算时间的步骤)
  8. Hidden states ( h_t )(隐藏状态 ( h_t ))
  9. Sequential nature precludes parallelization(序列性质排除了并行化)
  10. Memory constraints limit batching across examples(内存限制了跨样本的批处理)
  11. Factorization tricks(分解技巧)
  12. Conditional computation(条件计算)
  13. Sequential computation constraint(序列计算的限制)
  14. Modeling dependencies without regard to their distance(无需考虑距离的依赖关系建模)
  15. Eschewing recurrence(避开循环)
  16. Global dependencies(全局依赖)
  17. Parallelization(并行化)
  18. Translation quality(翻译质量)
  19. Compelling sequence modeling(引人注目的序列建模)
  20. Transduction models(转换模型)
  21. Inherently sequential nature(本质上的序列特性)
  22. Batching across examples(跨样本批处理)
  23. Computational efficiency(计算效率)
  24. Factorization tricks and conditional computation(分解技巧和条件计算)

2 Background

The goal of reducing sequential computation also forms the foundation of the Extended Neural GPU
[16], ByteNet [18] and ConvS2S [9], all of which use convolutional neural networks as basic building
block, computing hidden representations in parallel for all input and output positions.
In these models,
the number of operations required to relate signals from two arbitrary input or output positions grows
in the distance between positions, linearly for ConvS2S and logarithmically for ByteNet.
This makes
it more difficult to learn dependencies between distant positions [12].

减少序列计算的目标也构成了扩展神经 GPU(Extended Neural GPU)[16]、ByteNet [18] 和 ConvS2S [9] 的基础,这些模型都使用卷积神经网络作为基本构建块,为所有输入和输出位置并行计算隐藏表示。在这些模型中,将来自两个任意输入或输出位置的信号关联起来所需的操作数量,随着位置之间的距离而增长,对于 ConvS2S 来说是线性增长的,对于 ByteNet 则是对数增长的。这使得学习远距离位置之间的依赖关系更加困难 [12]。

这里说RNN网络的几个杰出代表Extended Neural GPU,ByteNet ,ConvS2S 也尝试实现并行运算,但是有个问题运算复杂度随着句子长度增加而增加,线性或者对数增加。

In the Transformer this is reduced to a constant number of operations, albeit at the cost of reduced effective resolution due to averaging attention-weighted positions, an effect we counteract with Multi-Head Attention as described in section 3.2.
在 Transformer 中,这被减少为恒定数量的操作,尽管代价是由于对注意力加权的位置进行平均,导致有效分辨率降低,我们通过第 3.2 节中描述的多头注意力来抵消这一影响。

transformer解决这个问题,运输量固定。运算量固定是通过注意力加权平均实现的,副作用是降低了有效的分辨率,这种副作用又通过多头注意力机制抵消

Self-attention, sometimes called intra-attention is an attention mechanism relating different positions
of a single sequence in order to compute a representation of the sequence.
Self-attention has been
used successfully in a variety of tasks including reading comprehension, abstractive summarization,
textual entailment and learning task-independent sentence representations [4, 27, 28, 22].

自注意力(有时称为内注意力)是一种注意力机制,用于关联单个序列中不同的位置,以计算该序列的表示。自注意力已成功应用于各种任务,包括阅读理解、抽象摘要、文本蕴含和学习任务无关的句子表示 [4, 27, 28, 22]。

再说一下注意力机制,是描述一个句子中每个单词和其它单词关系的机制

End-to-end memory networks are based on a recurrent attention mechanism instead of sequencealigned recurrence and have been shown to perform well on simple-language question answering and
language modeling tasks [34].

端到端记忆网络基于循环注意力机制,而不是序列对齐的循环,并已被证明在简单语言的问答和语言建模任务中表现良好 [34]。

这里夸一下End-to-end memory networks,但是这个是基于recurrent attention mechanism而不是sequencealigned recurrence。

To the best of our knowledge, however, the Transformer is the first transduction model relying
entirely on self-attention to compute representations of its input and output without using sequencealigned RNNs or convolution.
In the following sections, we will describe the Transformer, motivate
self-attention and discuss its advantages over models such as [17, 18] and [9].

然而,据我们所知,Transformer 是第一个完全依赖自注意力来计算其输入和输出表示的转换模型,而不使用序列对齐的循环神经网络或卷积。在接下来的章节中,我们将描述 Transformer,阐述自注意力的动机,并讨论其相对于 [17, 18] 和 [9] 等模型的优势。

这里再总结一下,Transformer 是第一个完全依赖自注意力来计算其输入和输出表示的转换模型。下面开始进入主题。
  1. Extended Neural GPU(扩展神经 GPU,论文之一)
  2. ByteNet(论文之一)
  3. ConvS2S(卷积序列到序列模型,论文之一)
  4. 基本构建块(Basic Building Block)
  5. 隐藏表示(Hidden Representations)
  6. 并行计算(Computing in Parallel)
  7. 将信号关联起来(Relate Signals)
  8. 任意输入或输出位置(Arbitrary Input or Output Positions)
  9. 操作数量(Number of Operations)
  10. 线性增长(Linearly)
  11. 对数增长(Logarithmically)
  12. 远距离位置之间的依赖关系(Dependencies Between Distant Positions)
  13. 有效分辨率(Effective Resolution)
  14. 注意力加权的位置取平均(Averaging Attention-Weighted Positions)
  15. 多头注意力(Multi-Head Attention)
  16. 自注意力 / 内注意力(Self-Attention / Intra-Attention)
  17. 注意力机制(Attention Mechanism)
  18. 阅读理解(Reading Comprehension)
  19. 抽象摘要(Abstractive Summarization)
  20. 文本蕴含(Textual Entailment)
  21. 任务无关的句子表示(Task-Independent Sentence Representations)
  22. 端到端记忆网络(End-to-End Memory Networks)
  23. 循环注意力机制(Recurrent Attention Mechanism)
  24. 序列对齐的循环(Sequence-Aligned Recurrence)
  25. 简单语言的问答(Simple-Language Question Answering)
  26. 语言建模任务(Language Modeling Tasks)
  27. 转换模型(Transduction Model)
  28. 序列对齐的循环神经网络(Sequence-Aligned RNNs)
  29. 卷积(Convolution)
  30. 阐述自注意力的动机(Motivate Self-Attention)
  31. 相对于其他模型的优势(Advantages Over Models Such As [17, 18] And [9])

3 Model Architecture

Most competitive neural sequence transduction models have an encoder-decoder structure [5, 2, 35].
Here, the encoder maps an input sequence of symbol representations (x1, …, xn) to a sequence
of continuous representations z = (z1, …, zn). Given z, the decoder then generates an output
sequence
(y1, …, ym) of symbols one element at a time. At each step the model is auto-regressive
[10], consuming the previously generated symbols as additional input when generating the next.
大多数具有竞争力的神经序列转换模型都有一个编码器-解码器结构 [5, 2, 35]。在这里,编码器将符号表示的输入序列 (x₁, …, xₙ) 映射到连续表示的序列 z = (z₁, …, zₙ)。给定 z,解码器然后一次生成一个符号的输出序列 (y₁, …, yₘ)。在每一步,模型都是自回归的 [10],在生成下一个符号时,使用先前生成的符号作为额外的输入。

先讲一下编码器-解码器架构的工作原理:x通过编码器变成z,然后z和y通过解码器变成y。z+y变成y的原理叫自回归。

The Transformer follows this overall architecture using stacked self-attention and point-wise, fully
connected layers for both the encoder and decoder, shown in the left and right halves of Figure 1,
respectively.
Transformer 遵循这种整体架构,对编码器和解码器都使用堆叠的自注意力逐点的全连接层,分别显示在图 1 的左半部分和右半部分。
在这里插入图片描述

这里说Transformer也是遵守编码器解码器这种架构的,上图是架构图Transformer的图。这里注意Outputs就是上面提到的y,shifted right就是上文中的“先前生成的符号”
  1. 神经序列转换模型(Neural Sequence Transduction Models)
  2. 编码器-解码器结构(Encoder-Decoder Structure)
  3. 符号表示(Symbol Representations)
  4. 连续表示(Continuous Representations)
  5. 自回归(Auto-Regressive)
  6. 使用先前生成的符号作为额外的输入(Consuming the Previously Generated Symbols as Additional Input)
  7. 堆叠的自注意力(Stacked Self-Attention)
  8. 逐点的全连接层(Point-Wise, Fully Connected Layers)

3.1 Encoder and Decoder Stacks

Encoder: The encoder is composed of a stack of N = 6 identical layers. Each layer has two
sub-layers. The first is a multi-head self-attention mechanism, and the second is a simple, positionwise fully connected feed-forward network. We employ a residual connection [11] around each of the two sub-layers, followed by layer normalization [1]. That is, the output of each sub-layer is L a y e r N o r m ( x + S u b l a y e r ( x ) ) LayerNorm(x + Sublayer(x)) LayerNorm(x+Sublayer(x)), where S u b l a y e r ( x ) Sublayer(x) Sublayer(x) is the function implemented by the sub-layer itself. To facilitate these residual connections, all sub-layers in the model, as well as the embedding layers, produce outputs of dimension d model = 512 d_{\text{model}} = 512 dmodel=512.

编码器:编码器由一个包含 N=6 个相同层的堆栈组成。每一层有两个子层。第一个是多头自注意力机制,第二个是一个简单的、逐位置的全连接前馈网络。我们在每个子层周围使用残差连接 [11],然后进行层归一化 [1]。也就是说,每个子层的输出是 LayerNorm(x + Sublayer(x)),其中 Sublayer(x) 是子层自身实现的函数。为了方便这些残差连接,模型中的所有子层以及嵌入层,都产生维度为 ( d_{\text{model}} = 512 ) 的输出。

这里是核心,讲transformers的编码器,一共6层,每层又两个子层,分别是多头自注意力层和全连接网络,
所有层都是通过残差和归一化LayerNorm(x + Sublayer(x))连接。也可以认为是一个小网络。

Decoder: The decoder is also composed of a stack of N = 6 identical layers. In addition to the two
sub-layers in each encoder layer, the decoder inserts a third sub-layer, which performs multi-head
attention over the output of the encoder stack
. Similar to the encoder, we employ residual connections
around each of the sub-layers, followed by layer normalization. We also modify the self-attention
sub-layer in the decoder stack to prevent positions from attending to subsequent positions.
This
masking, combined with fact that the output embeddings are offset by one position, ensures that the
predictions for position i can depend only on the known outputs at positions less than i.
解码器:解码器也由一个包含 N=6 个相同层的堆栈组成。除了每个编码器层中的两个子层外,解码器还插入了第三个子层,它对编码器堆栈的输出执行多头注意力。与编码器类似,我们在每个子层周围使用残差连接,然后进行层归一化。我们还修改了解码器堆栈中的自注意力子层,以防止当前位置关注后续的位置。这种掩码机制,加上输出嵌入偏移一个位置的事实,确保了位置 ( i ) 的预测只能依赖于位置小于 ( i ) 的已知输出。

解码器也是6层,每层有3个子层。两个多头注意力层,一个全连接层。第一个多头注意力层处理输出自回归的数据。第二个多头注意力层接收编码器输出(kv输入)及上一层输出的输出(q输入),第二层的残差连接也用上一层的输出,第三层是全连接层。
  1. 编码器和解码器堆栈(Encoder and Decoder Stacks)
  2. 子层(Sub-layers)
  3. 多头自注意力机制(Multi-Head Self-Attention Mechanism)
  4. 逐位置的全连接前馈网络(Position-Wise Fully Connected Feed-Forward Network)
  5. 残差连接(Residual Connection)
  6. 层归一化(Layer Normalization)
  7. LayerNorm(x + Sublayer(x)):一种计算方式,表示对输入和子层输出之和进行层归一化。
  8. Sublayer(x)(子层函数):子层本身实现的函数。
  9. 嵌入层(Embedding Layers)
  10. 维度 d model = 512 d_{\text{model}} = 512 dmodel=512(Dimension d model = 512 d_{\text{model}} = 512 dmodel=512):模型中向量表示的大小。
  11. 掩码机制(Masking)
  12. 输出嵌入偏移一个位置(Output Embeddings Are Offset by One Position)
  13. 防止当前位置关注后续的位置(Prevent Positions from Attending to Subsequent Positions)
  14. 位置 i i i 的预测只能依赖于位置小于 i i i的已知输出(Predictions for Position i i i Can Depend Only on Known Outputs at Positions Less Than i i i

3.2 Attention

An attention function can be described as mapping a query and a set of key-value pairs to an output,
where the query, keys, values, and output are all vectors. The output is computed as a weighted sum of the values, where the weight assigned to each value is computed by a compatibility function of the query with the corresponding key.
一个注意力函数可以被描述为将一个查询(query)和一组键-值对(key-value pairs)映射到一个输出,其中查询、键、值和输出都是向量(vectors)。输出被计算为值的加权和(weighted sum),其中分配给每个值的权重是通过查询与对应键的兼容性函数计算得到的。
在这里插入图片描述
Figure 2: (left) Scaled Dot-Product Attention. (right) Multi-Head Attention consists of several
attention layers running in parallel.
图2:(左)缩放点积注意力。(右)多头注意力由多个并行运行的注意力层组成。

  1. 注意力函数(Attention Function)
  2. 查询(query)(Query)
  3. 键-值对(Key-Value Pairs)
  4. 向量(Vectors)
  5. 加权和(Weighted Sum)
  6. 兼容性函数(Compatibility Function)
  7. 缩放点积注意力(Scaled Dot-Product Attention)
  8. 多头注意力(Multi-Head Attention)
  9. 并行运行的注意力层(Attention Layers Running in Parallel)

3.2.1 Scaled Dot-Product Attention

We call our particular attention “Scaled Dot-Product Attention” (Figure 2). The input consists of queries and keys of dimension d k d_k dk, and values of dimension d v d_v dv. We compute the dot products of the query with all keys, divide each by d k \sqrt{d_k} dk , and apply a softmax function to obtain the weights on the values.
我们将我们特定的注意力称为缩放点积注意力(见图2)。输入由维度为 d k d_k dk查询(queries)键(keys)以及维度为 d v d_v dv值(values)组成。我们计算查询与所有键的点积,然后除以 d k \sqrt{d_k} dk ,并应用softmax 函数来获得对值的权重。
In practice, we compute the attention function on a set of queries simultaneously, packed together into a matrix Q Q Q. The keys and values are also packed together into matrices K K K and V V V. We compute the matrix of outputs as:
在实践中,我们同时对一组查询计算注意力函数,将它们打包成一个矩阵 Q Q Q。键和值也分别打包成矩阵 K K K V V V。我们将输出的矩阵计算为:
Attention ( Q , K , V ) = softmax ( Q K ⊤ d k ) V ( 1 ) \text{Attention}(Q, K, V) = \text{softmax}\left( \dfrac{Q K^\top}{\sqrt{d_k}} \right) V \quad (1) Attention(Q,K,V)=softmax(dk QK)V(1)

The two most commonly used attention functions are additive attention [2], and dot-product (multiplicative) attention. Dot-product attention is identical to our algorithm, except for the scaling factor of 1 d k \dfrac{1}{\sqrt{d_k}} dk 1. Additive attention computes the compatibility function using a feed-forward network with a single hidden layer. While the two are similar in theoretical complexity, dot-product attention is much faster and more space-efficient in practice, since it can be implemented using highly optimized matrix multiplication code.
最常用的两种注意力函数是加性注意力 [2] 和点积(乘性)注意力。点积注意力与我们的算法相同,除了缩放因子 1 d k \dfrac{1}{\sqrt{d_k}} dk 1。加性注意力使用具有单个隐藏层的前馈网络来计算兼容性函数。虽然两者在理论复杂度上相似,但在实践中,点积注意力速度更快、空间效率更高,因为它可以使用高度优化的矩阵乘法代码实现。

这里将Transformers为什么选点积注意力,并说明对点积注意力改进,缩放因子改成平方根

While for small values of d k d_k dkthe two mechanisms perform similarly, additive attention outperforms dot product attention without scaling for larger values of d k d_k dk [3]. We suspect that for large values of d k d_k dk, the dot products grow large in magnitude, pushing the softmax function into regions where it has extremely small gradients. To counteract this effect, we scale the dot products by 1 d k \dfrac{1}{\sqrt{d_k}} dk 1.
虽然在较小的 d k d_k dk值下,这两种机制表现相似,但对于较大的 d k d_k dk值,未缩放的点积注意力的表现不如加性注意力 [3]。我们怀疑,对于较大的 d k d_k dk值,点积的幅值变得很大,使得softmax函数推向梯度极小的区域 4 ^4 4。为了解决这一问题,我们将点积除以 d k \sqrt{d_k} dk 进行缩放。

这里讲为什么要优化缩放因子
  1. 缩放点积注意力(Scaled Dot-Product Attention)
  2. 查询(queries)
  3. 键(keys)
  4. 值(values)
  5. 维度 d k d_k dk d v d_v dv(Dimension d k d_k dk, d v d_v dv
  6. 点积(Dot Product)
  7. softmax 函数
  8. 兼容性函数(Compatibility Function)
  9. 矩阵 Q , K , V Q, K, V Q,K,V
  10. Q K ⊤ Q K^\top QK(矩阵乘法, K ⊤ K^\top K K的转置)
  11. 加性注意力(Additive Attention)
  12. 乘性注意力(Multiplicative Attention)
  13. 前馈网络(Feed-Forward Network)
  14. 单个隐藏层(Single Hidden Layer)
  15. 理论复杂度(Theoretical Complexity)
  16. 高度优化的矩阵乘法代码(Highly Optimized Matrix Multiplication Code)
  17. 未缩放的点积注意力(Dot Product Attention without Scaling)
  18. softmax 函数的梯度极小区域(Regions Where the Softmax Function Has Extremely Small Gradients)
  19. 将点积除以 d k \sqrt{d_k} dk 进行缩放(Scale the Dot Products by Dividing by d k \sqrt{d_k} dk

3.2.2 Multi-Head Attention

Instead of performing a single attention function with d m o d e l d_{model} dmodel-dimensional keys, values, and queries, we found it beneficial to linearly project the queries, keys, and values h h h times with different, learned linear projections to dimensions d k d_k dk, d k d_k dk, and d v d_v dv, respectively. On each of these projected versions of queries, keys, and values, we then perform the attention function in parallel, yielding d v d_v dv-dimensional output values. These are concatenated and once again projected, resulting in the final values, as depicted in Figure 2.
与其使用具有 d m o d e l d_{model} dmodel维度的键、值和查询执行单一的注意力函数,我们发现使用不同的、可学习的线性映射将查询、键和值分别线性投影 h h h次到 d k d_k dk d k d_k dk d v d_v dv维度会更有益。在这些投影后的查询、键和值上,我们然后并行地执行注意力函数,产生 d v d_v dv维度的输出值。这些输出被连接起来并再次投影,得到最终的值,如图2所示。

Multi-head attention allows the model to jointly attend to information from different representation subspaces at different positions. With a single attention head, averaging inhibits this ability.
多头注意力允许模型在不同位置的不同表示子空间上共同关注信息。使用单个注意力头,取平均会抑制这种能力。

MultiHead ( Q , K , V ) = Concat ( head 1 , … , head h ) W O \text{MultiHead}(Q, K, V) = \text{Concat}(\text{head}_1, \ldots, \text{head}_h) W^O MultiHead(Q,K,V)=Concat(head1,,headh)WO

where head i = Attention ( Q W i Q , K W i K , V W i V ) \text{head}_i = \text{Attention}(Q W^Q_i, K W^K_i, V W^V_i) headi=Attention(QWiQ,KWiK,VWiV).
其中 head i = Attention ( Q W i Q , K W i K , V W i V ) \text{head}_i = \text{Attention}(Q W^Q_i, K W^K_i, V W^V_i) headi=Attention(QWiQ,KWiK,VWiV)

The projections are parameter matrices W i Q ∈ R d m o d e l × d k W^Q_i \in \mathbb{R}^{d_{model} \times d_k} WiQRdmodel×dk, W i K ∈ R d m o d e l × d k W^K_i \in \mathbb{R}^{d_{model} \times d_k} WiKRdmodel×dk, W i V ∈ R d m o d e l × d v W^V_i \in \mathbb{R}^{d_{model} \times d_v} WiVRdmodel×dv, and W O ∈ R h d v × d m o d e l W^O \in \mathbb{R}^{h d_v \times d_{model}} WORhdv×dmodel.
这些投影是参数矩阵 W i Q ∈ R d m o d e l × d k W^Q_i \in \mathbb{R}^{d_{model} \times d_k} WiQRdmodel×dk W i K ∈ R d m o d e l × d k W^K_i \in \mathbb{R}^{d_{model} \times d_k} WiKRdmodel×dk W i V ∈ R d m o d e l × d v W^V_i \in \mathbb{R}^{d_{model} \times d_v} WiVRdmodel×dv W O ∈ R h d v × d m o d e l W^O \in \mathbb{R}^{h d_v \times d_{model}} WORhdv×dmodel

In this work, we employ h = 8 h = 8 h=8 parallel attention layers, or heads. For each of these, we use d k = d v = d m o d e l / h = 64 d_k = d_v = d_{model} / h = 64 dk=dv=dmodel/h=64. Due to the reduced dimension of each head, the total computational cost is similar to that of single-head attention with full dimensionality.
在这项工作中,我们采用 h = 8 h = 8 h=8个并行的注意力层,或称为注意力头。对于每个头,我们使用 d k = d v = d m o d e l / h = 64 d_k = d_v = d_{model} / h = 64 dk=dv=dmodel/h=64。由于每个头的维度降低,总的计算成本与具有完整维度的单头注意力相似。

  1. 线性投影(Linearly Project)
  2. 可学习的线性映射(Learned Linear Projections)
  3. d k d_k dk d v d_v dv维度(Dimensions d k d_k dk, d v d_v dv
  4. 并行地执行注意力函数(Perform the Attention Function in Parallel)
  5. d v d_v dv维度的输出值 d v d_v dv-dimensional Output Values)
  6. 连接并再次投影(Concatenated and Once Again Projected)
  7. 表示子空间(Representation Subspaces)
  8. 取平均会抑制这种能力(Averaging Inhibits This)
  9. 参数矩阵(Parameter Matrices)
  10. 并行的注意力层或头(Parallel Attention Layers or Heads)
  11. 维度降低(Reduced Dimension)
  12. 总的计算成本(Total Computational Cost)
  13. 具有完整维度的单头注意力(Single-Head Attention with Full Dimensionality)

3.2.3 Applications of Attention in our Model

The Transformer uses multi-head attention in three different ways:
Transformer 在三种不同的方式中使用多头注意力:

• In “encoder-decoder attention” layers, the queries come from the previous decoder layer,
and the memory keys and values come from the output of the encoder. This allows every
position in the decoder to attend over all positions in the input sequence. This mimics the
typical encoder-decoder attention mechanisms in sequence-to-sequence models such as
[38, 2, 9].

  • 编码器-解码器注意力层中,查询(queries)来自于前一层的解码器,记忆的键(keys)和值(values)来自于编码器的输出。这使得解码器中的每个位置都能关注输入序列中的所有位置。这模拟了序列到序列模型(如 [38, 2, 9])中典型的编码器-解码器注意力机制。
    • The encoder contains self-attention layers. In a self-attention layer all of the keys, values
    and queries come from the same place, in this case, the output of the previous layer in the
    encoder. Each position in the encoder can attend to all positions in the previous layer of the
    encoder.
  • 编码器包含自注意力层。在自注意力层中,所有的键、值和查询都来自于同一地方,在这种情况下,是编码器中前一层的输出。编码器中的每个位置都可以关注编码器前一层的所有位置。
    • Similarly, self-attention layers in the decoder allow each position in the decoder to attend to
    all positions in the decoder up to and including that position. We need to prevent leftward
    information flow in the decoder to preserve the auto-regressive property. We implement this
    inside of scaled dot-product attention by masking out (setting to −∞) all values in the input
    of the softmax which correspond to illegal connections. See Figure 2.
  • 类似地,解码器中的自注意力层允许解码器中每个位置关注解码器中直到并包括该位置的所有位置。我们需要防止解码器中的左向信息流动以保持自回归属性。我们在缩放点积注意力中通过对 softmax 输入中对应非法连接的所有值进行掩码(设为 − ∞ -\infty )来实现这一点。参见图2。
  1. 输入序列(Input Sequence)
  2. 关注(Attend)
  3. 模拟(Mimic)
  4. 编码器-解码器注意力机制(Encoder-Decoder Attention Mechanisms)
  5. 序列到序列模型(Sequence-to-Sequence Models)
  6. 自注意力层(Self-Attention Layers)
  7. 左向信息流动(Leftward Information Flow)
  8. 自回归属性(Auto-Regressive Property)
  9. 缩放点积注意力(Scaled Dot-Product Attention)
  10. 掩码(Masking Out)
  11. 非法连接(Illegal Connections)
  12. 设为 − ∞ -\infty (Setting to − ∞ -\infty
  13. softmax 输入(Input of the Softmax)

3.3 Position-wise Feed-Forward Networks

In addition to attention sub-layers, each of the layers in our encoder and decoder contains a fully connected feed-forward network, which is applied to each position separately and identically. This consists of two linear transformations with a ReLU activation in between.
除了注意力子层外,我们的编码器和解码器中的每一层都包含一个全连接的前馈网络,分别且相同地应用于每个位置。这由两个线性变换组成,中间有一个 ReLU 激活函数。
FFN ( x ) = max ⁡ ( 0 , x W 1 + b 1 ) W 2 + b 2 ( 2 ) \text{FFN}(x) = \max(0, x W_1 + b_1) W_2 + b_2 \quad (2) FFN(x)=max(0,xW1+b1)W2+b2(2)

While the linear transformations are the same across different positions, they use different parameters from layer to layer. Another way of describing this is as two convolutions with kernel size 1.
虽然线性变换在不同的位置是相同的,但它们在不同的层中使用不同的参数。另一种描述方式是,这是两个核大小为 1 的卷积。

The dimensionality of input and output is d model = 512 d_{\text{model}} = 512 dmodel=512, and the inner-layer has dimensionality d ff = 2048 d_{\text{ff}} = 2048 dff=2048.
输入和输出的维度是 d model = 512 d_{\text{model}} = 512 dmodel=512,而内层的维度是 d ff = 2048 d_{\text{ff}} = 2048 dff=2048

  1. 逐位置前馈网络(Position-wise Feed-Forward Networks)
  2. 全连接的前馈网络(Fully Connected Feed-Forward Network)
  3. 分别且相同地应用于每个位置(Applied to Each Position Separately and Identically)
  4. 线性变换(Linear Transformations)
  5. ReLU 激活函数(ReLU Activation)
  6. FFN ( x ) = max ⁡ ( 0 , x W 1 + b 1 ) W 2 + b 2 \text{FFN}(x) = \max(0, x W_1 + b_1) W_2 + b_2 FFN(x)=max(0,xW1+b1)W2+b2:前馈网络的公式
  7. 不同层中使用不同的参数(Use Different Parameters from Layer to Layer)
  8. 核大小为 1 的卷积(Convolutions with Kernel Size 1)
  9. 维度 d model = 512 d_{\text{model}} = 512 dmodel=512(Dimensionality d model = 512 d_{\text{model}} = 512 dmodel=512
  10. 内层的维度 d ff = 2048 d_{\text{ff}} = 2048 dff=2048(Inner-layer Dimensionality d ff = 2048 d_{\text{ff}} = 2048 dff=2048

3.4 Embeddings and Softmax

Similarly to other sequence transduction models, we use learned embeddings to convert the input tokens and output tokens to vectors of dimension d model d_{\text{model}} dmodel. We also use the usual learned linear transformation and softmax function to convert the decoder output to predicted next-token probabilities. In our model, we share the same weight matrix between the two embedding layers and the pre-softmax linear transformation, similar to [30]. In the embedding layers, we multiply those weights by d model \sqrt{d_{\text{model}}} dmodel .
与其他序列转换模型类似,我们使用学习到的嵌入(embeddings)将输入标记和输出标记转换为维度为 d model d_{\text{model}} dmodel的向量。我们也使用常用的学习到的线性变换和 softmax 函数将解码器的输出转换为预测的下一个标记的概率。在我们的模型中,我们在两个嵌入层和 softmax 之前的线性变换之间共享相同的权重矩阵,类似于 [30]。在嵌入层中,我们将这些权重乘以 d model \sqrt{d_{\text{model}}} dmodel

  1. 嵌入(Embeddings)
  2. Softmax
  3. 序列转换模型(Sequence Transduction Models)
  4. 标记(Tokens)
  5. 向量(Vectors)
  6. 维度 d model d_{\text{model}} dmodel(Dimension d model d_{\text{model}} dmodel
  7. 线性变换(Linear Transformation)
  8. 权重矩阵(Weight Matrix)
  9. 卷积(Convolutions)
  10. 核大小 k k k(Kernel Size k k k
  11. 限制性自注意力(Restricted Self-Attention)
  12. 邻域大小 r r r(Size of the Neighborhood r r r
  13. 复杂度(Complexity)
  14. 序列操作数(Sequential Operations)
  15. 最大路径长度(Maximum Path Length)

3.5 Positional Encoding

Since our model contains no recurrence and no convolution, in order for the model to make use of the order of the sequence, we must inject some information about the relative or absolute position of the tokens in the sequence. To this end, we add “positional encodings” to the input embeddings at the bottoms of the encoder and decoder stacks. The positional encodings have the same dimension d model d_{\text{model}} dmodel as the embeddings, so that the two can be summed. There are many choices of positional encodings, learned and fixed [9].
由于我们的模型没有循环也没有卷积,为了让模型利用序列的顺序,我们必须为序列中标记的相对或绝对位置注入一些信息。为此,我们在编码器和解码器堆栈的底部将“位置编码”添加到输入嵌入中。位置编码与嵌入的维度 d model d_{\text{model}} dmodel 相同,因此两者可以相加。位置编码有很多选择,可以是学习得到的,也可以是固定的 [9]。

In this work, we use sine and cosine functions of different frequencies:
在这项工作中,我们使用不同频率的正弦和余弦函数:

P E ( pos ,   2 i ) = sin ⁡ ( pos 1000 0 2 i d model ) \mathrm{PE}_{(\text{pos},\,2i)} = \sin\left( \dfrac{\text{pos}}{10000^{\frac{2i}{d_{\text{model}}}}} \right) PE(pos,2i)=sin(10000dmodel2ipos)

P E ( pos ,   2 i + 1 ) = cos ⁡ ( pos 1000 0 2 i d model ) \mathrm{PE}_{(\text{pos},\,2i+1)} = \cos\left( \dfrac{\text{pos}}{10000^{\frac{2i}{d_{\text{model}}}}} \right) PE(pos,2i+1)=cos(10000dmodel2ipos)

where pos \text{pos} pos is the position and i i i is the dimension. That is, each dimension of the positional encoding corresponds to a sinusoid. The wavelengths form a geometric progression from 2 π 2\pi 2π to 10000 ⋅ 2 π 10000 \cdot 2\pi 100002π. We chose this function because we hypothesized it would allow the model to easily learn to attend by relative positions, since for any fixed offset k k k, P E pos + k \mathrm{PE}_{\text{pos}+k} PEpos+k can be represented as a linear function of P E pos \mathrm{PE}_{\text{pos}} PEpos.
其中 pos \text{pos} pos 是位置, i i i 是维度。也就是说,位置编码的每个维度对应一个正弦波。波长形成一个从 2 π 2\pi 2π 10000 ⋅ 2 π 10000 \cdot 2\pi 100002π 的等比数列。我们选择这个函数是因为我们假设它将允许模型轻松地通过相对位置进行关注,因为对于任何固定的偏移 k k k P E pos + k \mathrm{PE}_{\text{pos}+k} PEpos+k 可以表示为 P E pos \mathrm{PE}_{\text{pos}} PEpos 的线性函数。

We also experimented with using learned positional embeddings [9] instead, and found that the two versions produced nearly identical results (see Table 3 row (E)). We chose the sinusoidal version because it may allow the model to extrapolate to sequence lengths longer than the ones encountered during training.
我们也尝试了使用学习到的位置嵌入 [9],并发现两种版本产生了几乎相同的结果(见表 3 的第 (E) 行)。我们选择了正弦版,因为它可能允许模型对比训练中遇到的更长的序列长度进行外推。

  1. 位置编码(Positional Encoding)
  2. 序列的顺序(Order of the Sequence)
  3. 注入信息(Inject Information)
  4. 相对或绝对位置(Relative or Absolute Position)
  5. 标记(Tokens)
  6. 输入嵌入(Input Embeddings)
  7. 编码器和解码器堆栈的底部(Bottoms of the Encoder and Decoder Stacks)
  8. 维度 d model d_{\text{model}} dmodel(Dimension d model d_{\text{model}} dmodel
  9. 学习得到的或固定的(Learned and Fixed)
  10. 正弦和余弦函数(Sine and Cosine Functions)
  11. 不同频率(Different Frequencies)
  12. P E ( pos ,   2 i ) \mathrm{PE}_{(\text{pos},\,2i)} PE(pos,2i) P E ( pos ,   2 i + 1 ) \mathrm{PE}_{(\text{pos},\,2i+1)} PE(pos,2i+1):位置编码的公式
  13. 位置 pos \text{pos} pos(Position pos \text{pos} pos
  14. 维度 i i i(Dimension i i i
  15. 正弦波(Sinusoid)
  16. 波长形成等比数列(Wavelengths Form a Geometric Progression)
  17. 2 π 2\pi 2π 10000 ⋅ 2 π 10000 \cdot 2\pi 100002π:波长范围
  18. 假设(Hypothesize)
  19. 关注(Attend)
  20. 相对位置(Relative Positions)
  21. 固定偏移 k k k(Fixed Offset k k k
  22. 外推到更长的序列长度(Extrapolate to Sequence Lengths Longer)
  23. 训练中遇到的(Encountered During Training)

4 Why Self-Attention

In this section we compare various aspects of self-attention layers to the recurrent and convolutional layers commonly used for mapping one variable-length sequence of symbol representations ( x 1 , . . . , x n ) (x_1, ..., x_n) (x1,...,xn) to another sequence of equal length ( z 1 , . . . , z n ) (z_1, ..., z_n) (z1,...,zn), with x i , z i ∈ R d x_i, z_i \in \mathbb{R}^d xi,ziRd, such as a hidden layer in a typical sequence transduction encoder or decoder. Motivating our use of self-attention we consider three desiderata.
在本节中,我们比较了自注意力层与通常用于将一个可变长度的符号表示序列 ( x 1 , … , x n ) (x_1, \ldots, x_n) (x1,,xn) 映射到另一个相同长度的序列 ( z 1 , … , z n ) (z_1, \ldots, z_n) (z1,,zn) 的循环层和卷积层的各个方面,其中 x i , z i ∈ R d x_i, z_i \in \mathbb{R}^d xi,ziRd,例如典型的序列转换编码器或解码器中的隐藏层。为了激励我们使用自注意力机制,我们考虑了三个理想条件。

One is the total computational complexity per layer. Another is the amount of computation that can be parallelized, as measured by the minimum number of sequential operations required.
首先是每层的总计算复杂度。其次是可以并行化的计算量,用所需的最小顺序操作次数来衡量。

The third is the path length between long-range dependencies in the network. Learning long-range dependencies is a key challenge in many sequence transduction tasks. One key factor affecting the ability to learn such dependencies is the length of the paths forward and backward signals have to traverse in the network. The shorter these paths between any combination of positions in the input and output sequences, the easier it is to learn long-range dependencies [12]. Hence we also compare the maximum path length between any two input and output positions in networks composed of the different layer types.
第三个是网络中长距离依赖之间的路径长度。学习长距离依赖是许多序列转换任务中的关键挑战。影响学习此类依赖能力的一个关键因素是前向和后向信号在网络中必须遍历的路径长度。输入和输出序列中任何位置组合之间的这些路径越短,越容易学习长距离依赖 [12]。因此,我们还比较了由不同层类型组成的网络中任何两个输入和输出位置之间的最大路径长度。

Table 1: Maximum path lengths, per-layer complexity and minimum number of sequential operations for different layer types. n n n is the sequence length, d d d is the representation dimension, k k k is the kernel size of convolutions and r r r the size of the neighborhood in restricted self-attention.
表 1:不同层类型的每层复杂度、最小序列操作数和最大路径长度。其中, n n n 是序列长度, d d d 是表示维度, k k k 是卷积的核大小, r r r 是限制性自注意力中的邻域大小。

Layer TypeComplexity per LayerSequential OperationsMaximum Path Length
Self-Attention O ( n 2 ⋅ d ) O(n^2 \cdot d) O(n2d) O ( 1 ) O(1) O(1) O ( 1 ) O(1) O(1)
Recurrent O ( n ⋅ d 2 ) O(n \cdot d^2) O(nd2) O ( n ) O(n) O(n) O ( n ) O(n) O(n)
Convolutional O ( k ⋅ n ⋅ d 2 ) O(k \cdot n \cdot d^2) O(knd2) O ( 1 ) O(1) O(1) O ( log ⁡ k ( n ) ) O(\log_k(n)) O(logk(n))
Self-Attention (restricted) O ( r ⋅ n ⋅ d ) O(r \cdot n \cdot d) O(rnd) O ( 1 ) O(1) O(1) O ( n / r ) O(n/r) O(n/r)
层类型每层复杂度序列操作数最大路径长度
自注意力 O ( n 2 ⋅ d ) O(n^2 \cdot d) O(n2d) O ( 1 ) O(1) O(1) O ( 1 ) O(1) O(1)
循环(Recurrent) O ( n ⋅ d 2 ) O(n \cdot d^2) O(nd2) O ( n ) O(n) O(n) O ( n ) O(n) O(n)
卷积(Convolutional) O ( k ⋅ n ⋅ d 2 ) O(k \cdot n \cdot d^2) O(knd2) O ( 1 ) O(1) O(1) O ( log ⁡ k ( n ) ) O(\log_k(n)) O(logk(n))
自注意力(限制性) O ( r ⋅ n ⋅ d ) O(r \cdot n \cdot d) O(rnd) O ( 1 ) O(1) O(1) O ( n / r ) O(n/r) O(n/r)

As noted in Table 1, a self-attention layer connects all positions with a constant number of sequentially executed operations, whereas a recurrent layer requires O ( n ) O(n) O(n) sequential operations. In terms of computational complexity, self-attention layers are faster than recurrent layers when the sequence length n n n is smaller than the representation dimensionality d d d, which is most often the case with sentence representations used by state-of-the-art models in machine translation, such as word-piece [38] and byte-pair [31] representations. To improve computational performance for tasks involving very long sequences, self-attention could be restricted to considering only a neighborhood of size r r r in the input sequence centered around the respective output position. This would increase the maximum path length to O ( n / r ) O(n/r) O(n/r). We plan to investigate this approach further in future work.
如表 1 所示,自注意力层以常数次数的顺序执行操作连接所有位置,而循环层需要 O ( n ) O(n) O(n) 次顺序操作。在计算复杂度方面,当序列长度 n n n 小于表示维度 d d d 时,自注意力层比循环层更快,这在机器翻译的最新模型中使用的句子表示(如词片段 [38] 和字节对 [31] 表示)中经常是这种情况。为了提高涉及非常长序列的任务的计算性能,可以将自注意力机制限制为仅考虑以各自输出位置为中心的输入序列中大小为 r r r 的邻域。这将使最大路径长度增加到 O ( n / r ) O(n/r) O(n/r)。我们计划在未来的工作中进一步研究这种方法。

A single convolutional layer with kernel width k < n k < n k<n does not connect all pairs of input and output positions. Doing so requires a stack of O ( n / k ) O(n/k) O(n/k) convolutional layers in the case of contiguous kernels, or O ( log ⁡ k ( n ) ) O(\log_k(n)) O(logk(n)) in the case of dilated convolutions [18], increasing the length of the longest paths between any two positions in the network. Convolutional layers are generally more expensive than recurrent layers, by a factor of k k k. Separable convolutions [6], however, decrease the complexity considerably, to O ( k ⋅ n ⋅ d + n ⋅ d 2 ) O(k \cdot n \cdot d + n \cdot d^2) O(knd+nd2). Even with k = n k = n k=n, however, the complexity of a separable convolution is equal to the combination of a self-attention layer and a point-wise feed-forward layer, the approach we take in our model.
一个核宽度为 k < n k < n k<n 的单个卷积层并不能连接所有输入和输出位置对。在连续核的情况下,这需要堆叠 O ( n / k ) O(n/k) O(n/k) 个卷积层;在膨胀卷积 [18] 的情况下,需要堆叠 O ( log ⁡ k ( n ) ) O(\log_k(n)) O(logk(n)) 个卷积层,从而增加了网络中任何两个位置之间最长路径的长度。一般来说,卷积层比循环层的计算成本高,比例为 k k k。然而,深度可分离卷积 [6] 将复杂度显著降低到 O ( k ⋅ n ⋅ d + n ⋅ d 2 ) O(k \cdot n \cdot d + n \cdot d^2) O(knd+nd2)。即使当 k = n k = n k=n 时,深度可分离卷积的复杂度也等于自注意力层和逐点前馈层的组合,这正是我们在模型中采用的方法。

As a side benefit, self-attention could yield more interpretable models. We inspect attention distributions from our models and present and discuss examples in the appendix. Not only do individual attention heads clearly learn to perform different tasks, many appear to exhibit behavior related to the syntactic and semantic structure of the sentences.
作为附带的好处,自注意力机制可以产生更具可解释性的模型。我们检查了模型的注意力分布,并在附录中呈现和讨论了示例。不仅各个注意力头清晰地学习执行不同的任务,许多还表现出与句子的句法和语义结构相关的行为。

  1. 符号表示序列(Sequence of Symbol Representations)
  2. 计算复杂度(Computational Complexity)
  3. 并行化(Parallelization)
  4. 顺序操作次数(Number of Sequential Operations)
  5. 路径长度(Path Length)
  6. 长距离依赖(Long-Range Dependencies)
  7. 前向和后向信号(Forward and Backward Signals)
  8. 表示维度 d d d(Representation Dimensionality d d d
  9. 邻域大小 r r r(Neighborhood Size r r r
  10. 核宽度 k k k(Kernel Width k k k
  11. 连续核(Contiguous Kernels)
  12. 膨胀卷积(Dilated Convolutions)
  13. 深度可分离卷积(Separable Convolutions)
  14. 逐点前馈层(Point-Wise Feed-Forward Layer)
  15. 注意力头(Attention Heads)
  16. 句法和语义结构(Syntactic and Semantic Structure)

5 Training

This section describes the training regime for our models.
本节描述了我们模型的训练方案。

5.1 Training Data and Batching

We trained on the standard WMT 2014 English-German dataset consisting of about 4.5 million
sentence pairs. Sentences were encoded using byte-pair encoding [3], which has a shared sourcetarget vocabulary of about 37000 tokens. For English-French, we used the significantly larger WMT
2014 English-French dataset consisting of 36M sentences and split tokens into a 32000 word-piece
vocabulary [38]. Sentence pairs were batched together by approximate sequence length. Each training
batch contained a set of sentence pairs containing approximately 25000 source tokens and 25000
target tokens.
我们在标准的 WMT 2014 英语-德语数据集上进行了训练,该数据集包含约 450 万个句子对。句子使用字节对编码(byte-pair encoding)[3] 进行编码,具有大约 37000 个共享的源-目标词汇表。对于英语-法语,我们使用了更大规模的 WMT 2014 英语-法语数据集,包含 3600 万个句子,并将标记划分为一个由 32000 个词片(word-piece)组成的词汇表 [38]。句子对按照近似的序列长度进行分批处理。每个训练批次包含一组句子对,包含大约 25000 个源标记和 25000 个目标标记。

  1. 训练方案(Training Regime)
  2. 字节对编码(Byte-Pair Encoding)
  3. 共享的源-目标词汇表(Shared Source-Target Vocabulary)
  4. 词片(Word-Piece)
  5. 近似的序列长度(Approximate Sequence Length)
  6. 批处理(Batching)
  7. 训练批次(Training Batch)
  8. 源标记和目标标记(Source Tokens and Target Tokens)

5.2 Hardware and Schedule

We trained our models on one machine with 8 NVIDIA P100 GPUs. For our base models using
the hyperparameters described throughout the paper, each training step took about 0.4 seconds. We
trained the base models for a total of 100,000 steps or 12 hours. For our big models,(described on the
bottom line of table 3), step time was 1.0 seconds. The big models were trained for 300,000 steps
(3.5 days).
我们在一台配备 8 块 NVIDIA P100 GPU 的机器上训练了我们的模型。对于使用本文所述超参数的基础模型,每个训练步骤大约耗时 0.4 秒。我们总共训练了基础模型 100,000 个步骤,即 12 小时。对于我们的大型模型(见表 3 底部),每个步骤耗时 1.0 秒。大型模型训练了 300,000 个步骤(3.5 天)。

  1. 硬件和训练计划(Hardware and Schedule)
  2. 基础模型(Base Models)
  3. 超参数(Hyperparameters)
  4. 训练步骤(Training Step)
  5. 大型模型(Big Models)

5.3 Optimizer

We used the Adam optimizer [20] with β 1 = 0.9 \beta_1 = 0.9 β1=0.9, β 2 = 0.98 \beta_2 = 0.98 β2=0.98, and ϵ = 1 0 − 9 \epsilon = 10^{-9} ϵ=109. We varied the learning rate over the course of training, according to the formula:
我们使用了 Adam 优化器 [20],其中 β 1 = 0.9 \beta_1 = 0.9 β1=0.9 β 2 = 0.98 \beta_2 = 0.98 β2=0.98,以及 ϵ = 1 0 − 9 \epsilon = 10^{-9} ϵ=109。我们在训练过程中根据以下公式调整了学习率:
lrate = d model − 0.5 ⋅ min ⁡ ( step_num − 0.5 ,  step_num ⋅ warmup_steps − 1.5 ) ( 3 ) \text{lrate} = d_{\text{model}}^{-0.5} \cdot \min\left( \text{step\_num}^{-0.5},\ \text{step\_num} \cdot \text{warmup\_steps}^{-1.5} \right) \quad (3) lrate=dmodel0.5min(step_num0.5, step_numwarmup_steps1.5)(3)

This corresponds to increasing the learning rate linearly for the first warmup_steps \text{warmup\_steps} warmup_steps training steps, and decreasing it thereafter proportionally to the inverse square root of the step number. We used warmup_steps = 4000 \text{warmup\_steps} = 4000 warmup_steps=4000.
这对应于在前 warmup_steps \text{warmup\_steps} warmup_steps 个训练步骤中线性增加学习率,此后按步数的平方根倒数成比例地减少学习率。我们使用了 warmup_steps = 4000 \text{warmup\_steps} = 4000 warmup_steps=4000

  1. Adam 优化器(Adam Optimizer)
  2. β 1 \beta_1 β1, β 2 \beta_2 β2:Adam 优化器的超参数
  3. ϵ \epsilon ϵ:Adam 优化器中的极小值,防止除以零
  4. 学习率(lrate)(Learning Rate)
  5. d model d_{\text{model}} dmodel:模型的维度(Model Dimension)
  6. step_num \text{step\_num} step_num:当前训练的步数(Current Training Step Number)
  7. warmup_steps \text{warmup\_steps} warmup_steps:学习率预热的步数(Number of Warmup Steps)
  8. min ⁡ \min min:取最小值函数(Minimum Function)
  9. 平方根倒数(Inverse Square Root)
  10. 线性增加学习率(Linearly Increasing Learning Rate)
  11. 成比例地减少(Decreasing Proportionally to)
  12. 预热步骤(Warmup Steps)

5.4 Regularization

We employ three types of regularization during training:
我们在训练过程中采用了三种类型的正则化:

Table 2: The Transformer achieves better BLEU scores than previous state-of-the-art models on the English-to-German and English-to-French newstest2014 tests at a fraction of the training cost.
表 2:在只需一小部分训练成本的情况下,Transformer 在英语到德语和英语到法语的 newstest2014 测试中获得了比之前最先进的模型更高的 BLEU 分数。

ModelBLEU EN-DEBLEU EN-FRTraining Cost (FLOPs) EN-DETraining Cost (FLOPs) EN-FR
ByteNet [18]23.75
Deep-Att + PosUnk [39]39.21.0 · 1020
GNMT + RL [38]24.639.922.3 · 10191.4 · 1020
ConvS2S [9]25.1640.469.6 · 10181.5 · 1020
MoE [32]26.0340.562.0 · 10191.2 · 1020
Deep-Att + PosUnk Ensemble [39]40.48.0 · 1020
GNMT + RL Ensemble [38]26.3041.161.8 · 10201.1 · 1021
ConvS2S Ensemble [9]26.3641.297.7 · 10191.2 · 1021
Transformer (base model)27.338.13.3 · 1018
Transformer (big)28.441.82.3 · 1019
模型BLEU训练成本(FLOPs)
EN-DEEN-FR
ByteNet [18]23.75
Deep-Att + PosUnk [39]39.2
GNMT + RL [38]24.639.92
ConvS2S [9]25.1640.46
MoE [32]26.0340.56
Deep-Att + PosUnk Ensemble [39]40.4
GNMT + RL Ensemble [38]26.3041.16
ConvS2S Ensemble [9]26.3641.29
Transformer(基础模型)27.338.1
Transformer(大型模型)28.441.8

Residual Dropout We apply dropout [33] to the output of each sub-layer, before it is added to the sub-layer input and normalized. In addition, we apply dropout to the sums of the embeddings and the positional encodings in both the encoder and decoder stacks. For the base model, we use a rate of
Pdrop = 0.1.
残差 Dropout 我们在每个子层的输出上应用了 dropout [33],在它被添加到子层的输入并进行归一化之前。此外,我们在编码器和解码器堆栈中,也对嵌入和位置编码的和应用了 dropout。对于基础模型,我们使用的 dropout 率为: P drop = 0.1 P_{\text{drop}} = 0.1 Pdrop=0.1

Label Smoothing During training, we employed label smoothing of value ϵls = 0.1 [36]. This hurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score.
标签平滑 在训练过程中,我们采用了值为 ϵ l s = 0.1 \epsilon_{ls} = 0.1 ϵls=0.1 的标签平滑 [36]。这会增加困惑度(perplexity),因为模型学会了更加不确定,但提升了准确率和 BLEU 分数。

  1. 正则化(Regularization)
  2. 残差 Dropout(Residual Dropout)
  3. 子层(Sub-layer)
  4. 归一化(Normalized)
  5. Dropout 率(Dropout Rate)
  6. P drop P_{\text{drop}} Pdrop:Dropout 概率
  7. 标签平滑(Label Smoothing)
  8. ϵ l s \epsilon_{ls} ϵls:标签平滑的值
  9. 困惑度(Perplexity)
  10. 准确率(Accuracy)
  11. BLEU 分数(BLEU Score)
  12. 训练成本(FLOPs)(Training Cost (FLOPs))
  13. 词片模型(Word-Piece Model)
  14. 字节对编码(Byte-Pair Encoding)

6 Results

6.1 Machine Translation

On the WMT 2014 English-to-German translation task, the big transformer model (Transformer (big)
in Table 2) outperforms the best previously reported models (including ensembles) by more than 2.0
BLEU, establishing a new state-of-the-art BLEU score of 28.4. The configuration of this model is
listed in the bottom line of Table 3. Training took 3.5 days on 8 P100 GPUs. Even our base model
surpasses all previously published models and ensembles, at a fraction of the training cost of any of
the competitive models.
在 WMT 2014 英语到德语的翻译任务中,大型 Transformer 模型(表 2 中的 Transformer (big))比之前报道的最佳模型(包括集成模型)高出超过 2.0 的 BLEU 分数,达到了新的最先进 BLEU 分数 28.4。该模型的配置列在表 3 的底部。训练在 8 块 P100 GPU 上花了 3.5 天。即使是我们的基础模型,其性能也超过了之前所有已发布的模型和集成模型,而训练成本只是任何竞争模型的一小部分。
On the WMT 2014 English-to-French translation task, our big model achieves a BLEU score of 41.0,
outperforming all of the previously published single models, at less than 1/4 the training cost of the
previous state-of-the-art model. The Transformer (big) model trained for English-to-French used
dropout rate Pdrop = 0.1, instead of 0.3.
在 WMT 2014 英语到法语的翻译任务中,我们的大型模型取得了 41.0 的 BLEU 分数,超越了之前所有已发布的单模型,训练成本不到之前最先进模型的四分之一。用于英语到法语训练的 Transformer(big)模型使用了 dropout 率 P drop = 0.1 P_{\text{drop}} = 0.1 Pdrop=0.1,而不是 0.3。
For the base models, we used a single model obtained by averaging the last 5 checkpoints, which
were written at 10-minute intervals. For the big models, we averaged the last 20 checkpoints. We
used beam search with a beam size of 4 and length penalty α = 0.6 [38]. These hyperparameters
were chosen after experimentation on the development set. We set the maximum output length during
inference to input length + 50, but terminate early when possible [38].
对于基础模型,我们使用了通过平均最后 5 个检查点获得的单个模型,这些检查点每间隔 10 分钟保存一次。对于大型模型,我们平均了最后 20 个检查点。我们使用了束宽为 4、长度惩罚 α = 0.6 \alpha = 0.6 α=0.6 的束搜索 [38]。这些超参数是在开发集上经过实验后选择的。我们在推理过程中将最大输出长度设置为输入长度加 50,但在可能的情况下尽早终止 [38]。
Table 2 summarizes our results and compares our translation quality and training costs to other model
architectures from the literature. We estimate the number of floating point operations used to train a
model by multiplying the training time, the number of GPUs used, and an estimate of the sustained
single-precision floating-point capacity of each GPU 5
表 2 总结了我们的结果,并将我们的翻译质量和训练成本与文献中的其他模型架构进行了比较。我们通过将训练时间、使用的 GPU 数量以及对每个 GPU 的持续单精度浮点容量的估计相乘,来估计训练模型所使用的浮点运算次数。

  1. WMT 2014:2014 年机器翻译研讨会(Workshop on Machine Translation 2014)
  2. 集成模型(Ensemble Models)
  3. 检查点(Checkpoints)
  4. 束搜索(Beam Search)
  5. 束宽(Beam Size)
  6. 长度惩罚 α \alpha α(Length Penalty α \alpha α
  7. 超参数(Hyperparameters)
  8. 推理(Inference)
  9. 浮点运算次数(Number of Floating Point Operations)
  10. 单精度浮点容量(Single-Precision Floating-Point Capacity)

6.2 Model Variations

To evaluate the importance of different components of the Transformer, we varied our base model
in different ways, measuring the change in performance on English-to-German translation on the development set, newstest2013. We used beam search as described in the previous section, but no
checkpoint averaging. We present these results in Table 3.
为了评估 Transformer 不同组件的重要性,我们以不同的方式修改了我们的基础模型,测量在开发集 newstest2013 上英语到德语翻译性能的变化。我们采用了上一节中描述的束搜索算法,但没有进行检查点平均。我们在表 3 中展示了这些结果。
Table 3: Variations on the Transformer architecture. Unlisted values are identical to those of the base
model. All metrics are on the English-to-German translation development set, newstest2013. Listed
perplexities are per-wordpiece, according to our byte-pair encoding, and should not be compared to
per-word perplexities.
表3: Transformer架构的变体。未列出的数值与基础模型相同。所有指标均基于英语到德语的翻译开发集,newstest2013。列出的困惑度是基于我们的字节对编码的每个字片(wordpiece),不应与每字困惑度进行比较。

层数 (N)模型维度 (dmodel)前馈神经网络维度 (dff)头数 (h)关键维度 (dk)值维度 (dv)丢弃率 (Pdrop)标签平滑 (ϵls)训练步数困惑度 (开发集)BLEU (开发集)参数量 ×10^6
NdmodeldffhdkdvPdropϵlstrain stepsPPL (dev)BLEU (dev)params ×10^6
----------------------------------------------------------------------------------------------------
base65122048864640.10.1100K4.9225.865
(A)651220488N/AN/A0.10.1100K
1512512864640.10.1100K5.2924.9
4512204841281280.10.1100K5.0025.5
1651220481632320.10.1100K4.9125.8
3251220483216160.10.1100K5.0125.4
(B)165122048864640.10.1100K5.1625.158
325122048864640.10.1100K5.0125.460
©25122048864640.10.1100K6.1123.736
45122048864640.10.1100K5.1925.350
85122048864640.10.1100K4.8825.580
2565122048832320.10.1100K5.7524.528
1024512204881281280.10.1100K4.6626.0168
10245122048864640.10.1100K5.1225.453
40965122048864640.10.1100K4.7526.290
(D)6512204886464N/Aϵls100K
65122048864640.00.1100K5.7724.6
65122048864640.20.1100K4.9525.5
65122048864640.10.0100K4.6725.3
65122048864640.10.2100K5.4725.7
(E)65122048864640.10.1100K4.9225.7
big61024409616N/AN/A0.30.1300K4.3326.4213

In Table 3 rows (A), we vary the number of attention heads and the attention key and value dimensions,
keeping the amount of computation constant, as described in Section 3.2.2. While single-head
attention is 0.9 BLEU worse than the best setting, quality also drops off with too many heads.
在表 3 的 (A) 行中,我们改变了注意力头的数量以及注意力键和值的维度,同时保持计算量不变,如第 3.2.2 节所述。虽然单头注意力比最佳设置低了 0.9 的 BLEU 分数,但过多的头数也会导致质量下降。

In Table 3 rows (B), we observe that reducing the attention key size dk hurts model quality. This
suggests that determining compatibility is not easy and that a more sophisticated compatibility
function than dot product may be beneficial. We further observe in rows © and (D) that, as expected,
bigger models are better, and dropout is very helpful in avoiding over-fitting. In row (E) we replace our
sinusoidal positional encoding with learned positional embeddings [9], and observe nearly identical
results to the base model.
在表 3 的 (B) 行中,我们观察到减少注意力键大小 (d_k) 会降低模型质量。这表明确定兼容性并不容易,使用比点积更复杂的兼容性函数可能是有益的。我们在 © 和 (D) 行中进一步观察到,正如预期的那样,更大的模型表现更好,而 dropout 在避免过拟合方面非常有帮助。在 (E) 行中,我们用学习的位置信嵌入 [9] 替换了我们的正弦位置编码,发现与基础模型的结果几乎相同。

6.3 English Constituency Parsing

To evaluate if the Transformer can generalize to other tasks we performed experiments on English
constituency parsing. This task presents specific challenges: the output is subject to strong structural
constraints and is significantly longer than the input. Furthermore, RNN sequence-to-sequence
models have not been able to attain state-of-the-art results in small-data regimes [37].
为了评估 Transformer 是否能够泛化到其他任务,我们在英语成分句法分析上进行了实验。该任务具有特定的挑战:输出受到强烈的结构约束,并且显著地长于输入。此外,RNN 序列到序列模型在小数据规模下无法达到最先进的成果 [37]。

We trained a 4-layer transformer with dmodel = 1024 on the Wall Street Journal (WSJ) portion of the
Penn Treebank [25], about 40K training sentences. We also trained it in a semi-supervised setting,
using the larger high-confidence and BerkleyParser corpora from with approximately 17M sentences
[37]. We used a vocabulary of 16K tokens for the WSJ only setting and a vocabulary of 32K tokens
for the semi-supervised setting.
我们在 Penn Treebank [25] 的华尔街日报(WSJ)部分(约 4 万个训练句子)上,训练了一个具有 d_model = 1024 的 4 层 Transformer。我们还在半监督的设置下进行了训练,使用了更大的高置信度和 BerkeleyParser 语料库,约有 1700 万个句子 [37]。对于仅 WSJ 的设置,我们使用了包含 1.6 万个词的词汇表;对于半监督设置,我们使用了包含 3.2 万个词的词汇表。

We performed only a small number of experiments to select the dropout, both attention and residual
(section 5.4), learning rates and beam size on the Section 22 development set, all other parameters
remained unchanged from the English-to-German base translation model. During inference, we increased the maximum output length to input length + 300. We used a beam size of 21 and α = 0.3
for both WSJ only and the semi-supervised setting.
我们仅在第 22 节开发集上进行了少量实验,以选择 dropout(包括注意力和残差,见第 5.4 节)、学习率和 beam size,其他所有参数都与英语到德语的基础翻译模型保持一致。在推理过程中,我们将最大输出长度增加到输入长度加 300。对于仅 WSJ 和半监督设置,我们都使用了 beam size 为 21,α = 0.3。

Table 4: The Transformer generalizes well to English constituency parsing (Results are on Section 23 of WSJ)
表4:Transformer在英语成分解析任务上表现良好(结果基于WSJ的第23节)

解析器训练方式WSJ 23 F1
ParserTrainingWSJ 23 F1
Vinyals & Kaiser et al. (2014) [37]WSJ only, discriminative88.3
Petrov et al. (2006) [29]WSJ only, discriminative90.4
Zhu et al. (2013) [40]WSJ only, discriminative90.4
Dyer et al. (2016) [8]WSJ only, discriminative91.7
Transformer (4 layers)WSJ only, discriminative91.3
Zhu et al. (2013) [40]semi-supervised91.3
Huang & Harper (2009) [14]semi-supervised91.3
McClosky et al. (2006) [26]semi-supervised92.1
Vinyals & Kaiser et al. (2014) [37]semi-supervised92.1
Transformer (4 layers)semi-supervised92.7
Luong et al. (2015) [23]multi-task93.0
Dyer et al. (2016) [8]generative93.3

Our results in Table 4 show that despite the lack of task-specific tuning our model performs surprisingly well, yielding better results than all previously reported models with the exception of the
Recurrent Neural Network Grammar [8].
表 4 中的结果显示,尽管缺乏针对任务的特定调优,我们的模型仍表现出令人惊讶的优秀表现,除了循环神经网络语法模型 [8] 外,优于所有先前报道的模型。
In contrast to RNN sequence-to-sequence models [37], the Transformer outperforms the BerkeleyParser [29] even when training only on the WSJ training set of 40K sentences.
与 RNN 序列到序列模型 [37] 相比,即使仅在 WSJ 的 4 万句训练集上训练,Transformer 也优于 BerkeleyParser [29]。

7 Conclusion

In this work, we presented the Transformer, the first sequence transduction model based entirely on
attention, replacing the recurrent layers most commonly used in encoder-decoder architectures with
multi-headed self-attention.
在这项工作中,我们提出了 Transformer,这是第一个完全基于注意力机制的序列转换模型,用多头自注意力取代了编码器-解码器架构中最常用的循环层。

For translation tasks, the Transformer can be trained significantly faster than architectures based
on recurrent or convolutional layers. On both WMT 2014 English-to-German and WMT 2014
English-to-French translation tasks, we achieve a new state of the art. In the former task our best
model outperforms even all previously reported ensembles.
对于翻译任务,Transformer 比基于循环或卷积层的架构具有显著更快的训练速度。在 WMT 2014 英译德和 WMT 2014 英译法的翻译任务中,我们取得了新的最先进成果。在前一个任务中,我们的最佳模型甚至超越了之前所有已报道的集成模型。

We are excited about the future of attention-based models and plan to apply them to other tasks. We
plan to extend the Transformer to problems involving input and output modalities other than text and
to investigate local, restricted attention mechanisms to efficiently handle large inputs and outputs
such as images, audio and video. Making generation less sequential is another research goals of ours.

我们对基于注意力的模型的未来感到兴奋,并计划将它们应用于其他任务。我们计划将 Transformer 扩展到涉及非文本的输入和输出模态的问题,并研究局部的、受限的注意力机制,以高效处理大规模的输入和输出,如图像、音频和视频。减少生成的序列性也是我们的研究目标之一。
The code we used to train and evaluate our models is available at https://github.com/
tensorflow/tensor2tensor.
我们用于训练和评估模型的代码已公开在:https://github.com/tensorflow/tensor2tensor。

Acknowledgements We are grateful to Nal Kalchbrenner and Stephan Gouws for their fruitful
comments, corrections and inspiration.
致谢 我们感谢 Nal Kalchbrenner 和 Stephan Gouws 的宝贵意见、修正和启发。

References 参考文献

[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.
Jimmy Lei Ba, Jamie Ryan Kiros, 和 Geoffrey E Hinton. 层归一化。arXiv 预印本 arXiv:1607.06450, 2016.

[2] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. CoRR, abs/1409.0473, 2014.
Dzmitry Bahdanau, Kyunghyun Cho, 和 Yoshua Bengio. 通过联合学习来对齐和翻译的神经机器翻译。CoRR, abs/1409.0473, 2014.

[3] Denny Britz, Anna Goldie, Minh-Thang Luong, and Quoc V. Le. Massive exploration of neural machine translation architectures. CoRR, abs/1703.03906, 2017.
Denny Britz, Anna Goldie, Minh-Thang Luong, 和 Quoc V. Le. 神经机器翻译架构的大规模探索。CoRR, abs/1703.03906, 2017.

[4] Jianpeng Cheng, Li Dong, and Mirella Lapata. Long short-term memory-networks for machine reading. arXiv preprint arXiv:1601.06733, 2016.
Jianpeng Cheng, Li Dong, 和 Mirella Lapata. 用于机器阅读的长短期记忆网络。arXiv 预印本 arXiv:1601.06733, 2016.

[5] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. CoRR, abs/1406.1078, 2014.
Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, 和 Yoshua Bengio. 使用 RNN 编码解码器进行统计机器翻译中的短语表示学习。CoRR, abs/1406.1078, 2014.

[6] Francois Chollet. Xception: Deep learning with depthwise separable convolutions. arXiv preprint arXiv:1610.02357, 2016.
Francois Chollet. Xception: 使用深度可分离卷积的深度学习。arXiv 预印本 arXiv:1610.02357, 2016.

[7] Junyoung Chung, Çaglar Gülçehre, Kyunghyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling. CoRR, abs/1412.3555, 2014.
Junyoung Chung, Çaglar Gülçehre, Kyunghyun Cho, 和 Yoshua Bengio. 门控循环神经网络在序列建模中的经验评估。CoRR, abs/1412.3555, 2014.

[8] Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. Recurrent neural network grammars. In Proc. of NAACL, 2016.
Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, 和 Noah A. Smith. 循环神经网络语法。在 NAACL 会议论文集,2016 年。

[9] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolutional sequence to sequence learning. arXiv preprint arXiv:1705.03122v2, 2017.
Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, 和 Yann N. Dauphin. 卷积序列到序列学习。arXiv 预印本 arXiv:1705.03122v2, 2017.

[10] Alex Graves. Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850, 2013.
Alex Graves. 使用循环神经网络生成序列。arXiv 预印本 arXiv:1308.0850, 2013.

[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, 和 Jian Sun. 用于图像识别的深度残差学习。在 IEEE 计算机视觉与模式识别会议论文集,页码 770–778,2016 年。

[12] Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, and Jürgen Schmidhuber. Gradient flow in recurrent nets: the difficulty of learning long-term dependencies, 2001.
Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, 和 Jürgen Schmidhuber. 循环网络中的梯度流动:学习长期依赖的困难,2001 年。

[13] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997.
Sepp Hochreiter 和 Jürgen Schmidhuber. 长短期记忆。神经计算,9(8):1735–1780, 1997.

[14] Zhongqiang Huang and Mary Harper. Self-training PCFG grammars with latent annotations across languages. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 832–841. ACL, August 2009.
Zhongqiang Huang 和 Mary Harper. 跨语言的基于隐性注释的 PCFG 语法自我训练。在 2009 年自然语言处理实证方法会议论文集,页码 832–841。ACL,2009 年 8 月。

[15] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring the limits of language modeling. arXiv preprint arXiv:1602.02410, 2016.
Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, 和 Yonghui Wu. 探索语言建模的极限。arXiv 预印本 arXiv:1602.02410, 2016.

[16] Łukasz Kaiser and Samy Bengio. Can active memory replace attention? In Advances in Neural Information Processing Systems, (NIPS), 2016.
Łukasz Kaiser 和 Samy Bengio. 活动记忆可以代替注意力吗?在神经信息处理系统进展 (NIPS) 中,2016 年。

[17] Łukasz Kaiser and Ilya Sutskever. Neural GPUs learn algorithms. In International Conference on Learning Representations (ICLR), 2016.
Łukasz Kaiser 和 Ilya Sutskever. 神经 GPU 学习算法。在国际学习表征会议 (ICLR) 中,2016 年。

[18] Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Koray Kavukcuoglu. Neural machine translation in linear time. arXiv preprint arXiv:1610.10099v2, 2017.
Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, 和 Koray Kavukcuoglu. 线性时间的神经机器翻译。arXiv 预印本 arXiv:1610.10099v2, 2017.

[19] Yoon Kim, Carl Denton, Luong Hoang, and Alexander M. Rush. Structured attention networks. In International Conference on Learning Representations, 2017.
Yoon Kim, Carl Denton, Luong Hoang, 和 Alexander M. Rush. 结构化注意力网络。在学习表征国际会议,2017 年。

[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015.
Diederik Kingma 和 Jimmy Ba. Adam: 一种随机优化方法。在 ICLR,2015 年。

[21] Oleksii Kuchaiev and Boris Ginsburg. Factorization tricks for LSTM networks. arXiv preprint arXiv:1703.10722, 2017.
Oleksii Kuchaiev 和 Boris Ginsburg. LSTM 网络的因子分解技巧。arXiv 预印本 arXiv:1703.10722, 2017.

[22] Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen Zhou, and Yoshua Bengio. A structured self-attentive sentence embedding. arXiv preprint arXiv:1703.03130, 2017.
Zhouhan Lin, Minwei Feng, Cicero Nogueira dos Santos, Mo Yu, Bing Xiang, Bowen Zhou, 和 Yoshua Bengio. 结构化自注意力句子嵌入。arXiv 预印本 arXiv:1703.03130, 2017.

[23] Minh-Thang Luong, Quoc V. Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. Multi-task sequence to sequence learning. arXiv preprint arXiv:1511.06114, 2015.
Minh-Thang Luong, Quoc V. Le, Ilya Sutskever, Oriol Vinyals, 和 Lukasz Kaiser. 多任务序列到序列学习。arXiv 预印本 arXiv:1511.06114, 2015.

[24] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025, 2015.
Minh-Thang Luong, Hieu Pham, 和 Christopher D Manning. 基于注意力的神经机器翻译的有效方法。arXiv 预印本 arXiv:1508.04025, 2015.

[25] Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. Building a large annotated corpus of english: The penn treebank. Computational linguistics, 19(2):313–330, 1993.
Mitchell P Marcus, Mary Ann Marcinkiewicz, 和 Beatrice Santorini. 构建一个大型英语标注语料库:宾州树库。计算语言学,19(2):313–330, 1993.

[26] David McClosky, Eugene Charniak, and Mark Johnson. Effective self-training for parsing. In Proceedings of the Human Language Technology Conference of the NAACL, Main Conference, pages 152–159. ACL, June 2006.
David McClosky, Eugene Charniak, 和 Mark Johnson. 有效的解析自我训练。在 NAACL 人类语言技术会议主会议论文集,页码 152–159。ACL,2006 年 6 月。

[27] Ankur Parikh, Oscar Täckström, Dipanjan Das, and Jakob Uszkoreit. A decomposable attention model. In Empirical Methods in Natural Language Processing, 2016.
Ankur Parikh, Oscar Täckström, Dipanjan Das, 和 Jakob Uszkoreit. 一种可分解的注意力模型。在自然语言处理实证方法中,2016 年。

[28] Romain Paulus, Caiming Xiong, and Richard Socher. A deep reinforced model for abstractive summarization. arXiv preprint arXiv:1705.04304, 2017.
Romain Paulus, Caiming Xiong, 和 Richard Socher. 用于抽象总结的深度强化模型。arXiv 预印本 arXiv:1705.04304, 2017.

[29] Slav Petrov, Leon Barrett, Romain Thibaux, and Dan Klein. Learning accurate, compact, and interpretable tree annotation. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL, pages 433–440. ACL, July 2006.
Slav Petrov, Leon Barrett, Romain Thibaux, 和 Dan Klein. 学习精确、紧凑和可解释的树注释。在第 21 届计算语言学国际会议和 ACL 第 44 届年会论文集,页码 433–440。ACL,2006 年 7 月。

[30] Ofir Press and Lior Wolf. Using the output embedding to improve language models. arXiv preprint arXiv:1608.05859, 2016.
Ofir Press 和 Lior Wolf. 使用输出嵌入来改进语言模型。arXiv 预印本 arXiv:1608.05859, 2016.

[31] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015.
Rico Sennrich, Barry Haddow, 和 Alexandra Birch. 使用子词单元的稀有词神经机器翻译。arXiv 预印本 arXiv:1508.07909, 2015.

[32] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017.
Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, 和 Jeff Dean. 极其庞大的神经网络:稀疏门控专家混合层。arXiv 预印本 arXiv:1701.06538, 2017.

[33] Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(1):1929–1958, 2014.
Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, 和 Ruslan Salakhutdinov. Dropout: 防止神经网络过拟合的简单方法。机器学习研究杂志,15(1):1929–1958, 2014.

[34] Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. End-to-end memory networks. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems 28, pages 2440–2448. Curran Associates, Inc., 2015.
Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, 和 Rob Fergus. 端到端记忆网络。在 C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama 和 R. Garnett 主编的神经信息处理系统第 28 卷进展中,页码 2440–2448. Curran Associates, Inc., 2015.

[35] Ilya Sutskever, Oriol Vinyals, and Quoc VV Le. Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems, pages 3104–3112, 2014.
Ilya Sutskever, Oriol Vinyals, 和 Quoc VV Le. 使用神经网络的序列到序列学习。在神经信息处理系统进展中,页码 3104–3112, 2014.

[36] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. CoRR, abs/1512.00567, 2015.
Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, 和 Zbigniew Wojna. 重新思考用于计算机视觉的 Inception 架构。CoRR, abs/1512.00567, 2015.

[37] Vinyals & Kaiser, Koo, Petrov, Sutskever, and Hinton. Grammar as a foreign language. In Advances in Neural Information Processing Systems, 2015.
Vinyals, Kaiser, Koo, Petrov, Sutskever, 和 Hinton. 像对待外语一样对待语法。在神经信息处理系统进展中,2015 年。

[38] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144, 2016.
Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey 等人。谷歌的神经机器翻译系统:弥合人类翻译与机器翻译之间的差距。arXiv 预印本 arXiv:1609.08144, 2016.

[39] Jie Zhou, Ying Cao, Xuguang Wang, Peng Li, and Wei Xu. Deep recurrent models with fast-forward connections for neural machine translation. CoRR, abs/1606.04199, 2016.
Jie Zhou, Ying Cao, Xuguang Wang, Peng Li, 和 Wei Xu. 用于神经机器翻译的快速前向连接深度循环模型。CoRR, abs/1606.04199, 2016.

[40] Muhua Zhu, Yue Zhang, Wenliang Chen, Min Zhang, and Jingbo Zhu. Fast and accurate shift-reduce constituent parsing. In Proceedings of the 51st Annual Meeting of the ACL (Volume 1: Long Papers), pages 434–443. ACL, August 2013.
Muhua Zhu, Yue Zhang, Wenliang Chen, Min Zhang, 和 Jingbo Zhu. 快速且精确的转移归约成分解析。在第 51 届 ACL 年会上 (卷 1: 长篇论文) 论文集,页码 434–443. ACL, August 2013.

Attention Visualizations

在这里插入图片描述
Figure 3: An example of the attention mechanism following long-distance dependencies in the
encoder self-attention in layer 5 of 6. Many of the attention heads attend to a distant dependency of
the verb ‘making’, completing the phrase ‘making…more difficult’. Attentions here shown only for
the word ‘making’. Different colors represent different heads. Best viewed in color.
图3:这是一个注意力机制的示例,在第6层中的第5层编码器自注意力中跟踪长距离依赖关系。许多注意力头关注动词“making”的远程依赖,完成了短语“making…more difficult”(使……更加困难)。这里的注意力仅针对词语“making”进行展示。不同的颜色代表不同的注意力头。最好以彩色观看。

在这里插入图片描述
Figure 4: Two attention heads, also in layer 5 of 6, apparently involved in anaphora resolution. Top:
Full attentions for head 5. Bottom: Isolated attentions from just the word ‘its’ for attention heads 5
and 6. Note that the attentions are very sharp for this word.
图4:两个注意力头,同样位于6层中的第5层,显然参与了指代消解。上图:注意力头5的完整注意力。下图:仅针对单词“its”的注意力头5和6。请注意,对于这个词,注意力非常集中。

在这里插入图片描述
Figure 5: Many of the attention heads exhibit behaviour that seems related to the structure of the
sentence. We give two such examples above, from two different heads from the encoder self-attention
at layer 5 of 6. The heads clearly learned to perform different tasks.

图5:许多注意力头展示出与句子结构相关的行为。我们在上面给出了两个这样的例子,来自第6层中的第5层编码器自注意力的两个不同的头。这些头显然学会了执行不同的任务。

评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值