2017 AttentionisallYouNeed

From GM-RKB
(Redirected from Attention Is All You Need)
Jump to navigation Jump to search

Subject Headings: Transformer Neural Network, Self-Attention Mechanism, Positional Encoding, Layer Normalization.

Notes

Cited By

2023

  • chat
    • The paper "Attention is All You Need" by Vaswani et al. introduced the Transformer model, which has since become the foundation for many state-of-the-art natural language processing models. Three key innovations presented in the paper are:
      • Self-Attention Mechanism: The most notable innovation introduced in the paper is the self-attention mechanism, which allows the model to weigh the importance of each input token relative to the others when making predictions. This mechanism enables the model to capture relationships and dependencies among the input tokens without relying on recurrent or convolutional layers. The self-attention mechanism is computationally efficient and can be easily parallelized, leading to faster training times compared to traditional RNN-based architectures.
      • Positional Encoding: Since the Transformer model does not have any inherent sense of the position of tokens in a sequence, the authors introduced positional encoding to inject information about the position of tokens in the input sequence. Positional encoding is added to the input embeddings before being processed by the self-attention layers, allowing the model to learn and use positional information when making predictions. The authors used a sinusoidal function to generate the positional encodings, ensuring that the encodings can be easily extended to varying sequence lengths.
      • Layer Normalization and Residual Connections: The Transformer architecture employs layer normalization and residual connections in both the encoder and decoder components. Layer normalization helps stabilize the training process and speeds up convergence by normalizing the activations of each layer. Residual connections, on the other hand, allow the model to learn more effectively by enabling the flow of information from earlier layers to later layers. These architectural features help the Transformer model to be trained effectively, even when it has a large number of layers.

2018b

2018c

Quotes

Abstract

The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder and decoder configuration. The best performing such models also connect the encoder and decoder through an attention mechanisms. We propose a novel, simple network architecture based solely on an attention mechanism, dispensing with recurrence and convolutions entirely. 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 single model with 165 million parameters, achieves 27.5 BLEU on English-to-German translation, improving over the existing best ensemble result by over 1 BLEU. On English-to-French translation, we outperform the previous single state-of-the-art with model by 0.7 BLEU, achieving a BLEU score of 41.1.

Notes about the Authors

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].

 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 $h_t$, as a function of the previous hidden state $h_{t-1}$ 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.

 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.

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 parallelization and can reach a new state-of-the-art in translation quality after being trained for as little as twelve hours on eight P100 GPUs.

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]. 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.

 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, 22, 23, 19].

 End-to-end memory networks are based on a recurrent attention mechanism instead of sequence aligned recurrence and have been shown to perform well on simple-language question answering and language modeling tasks [28].

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 sequence aligned RNNs or convolution.

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 [math]\displaystyle{ (x_1, \cdots , x_n) }[/math] to a sequence of continuous representations [math]\displaystyle{ \mathbf{z} = (z_1, \cdots, z_n) }[/math]. Given [math]\displaystyle{ \mathbf{z} }[/math], the decoder then generates an output sequence [math]\displaystyle{ (y_1, \cdots, y_m) }[/math] 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.

The Transformer follows this overall architecture using stacked self-attention and pointwise, fully connected layers for both the encoder and decoder, shown in the left and right halves of Figure 1, respectively.

3.1 Encoder and Decoder Stacks

Encoder: The encoder is composed of a stack of [math]\displaystyle{ N = 6 }[/math] identical layers. Each layer has two sub-layers. The first is a multi-head self-attention mechanism, and the second is a simple, position-wise 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 [math]\displaystyle{ LayerNorm(x + Sublayer(x)) }[/math], where [math]\displaystyle{ Sublayer(x) }[/math] 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 [math]\displaystyle{ d_{model} = 512 }[/math].

Decoder: The decoder is also composed of a stack of [math]\displaystyle{ N = 6 }[/math] 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 [math]\displaystyle{ i }[/math] can depend only on the known outputs at positions less than [math]\displaystyle{ i }[/math].

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.

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 [math]\displaystyle{ d_k }[/math], and values of dimension [math]\displaystyle{ d_v }[/math]. We compute the dot products of the query with all keys, divide each by [math]\displaystyle{ \sqrt{d_k} }[/math] , and apply a softmax function to obtain the weights on the values.

2017 AttentionisallYouNeed Fig2.png
Figure 2: (left) Scaled Dot-Product Attention. (right) Multi-Head Attention consists of several attention layers running in parallel.

In practice, we compute the attention function on a set of queries simultaneously, packed together into a matrix [math]\displaystyle{ Q }[/math]. The keys and values are also packed together into matrices [math]\displaystyle{ K }[/math] and [math]\displaystyle{ V }[/math]. We compute the matrix of outputs as:

[math]\displaystyle{ Attention(Q,K,V)=\mathrm{softmax}\left(\dfrac{QK_T}{\sqrt{d_k}}\right)V }[/math] (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 [math]\displaystyle{ 1/\sqrt{d_k} }[/math]. 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.

While for small values of [math]\displaystyle{ d_k }[/math] the two mechanisms perform similarly, additive attention outperforms dot product attention without scaling for larger values of [math]\displaystyle{ d_k }[/math] [3]. We suspect that for large values of [math]\displaystyle{ d_k }[/math], the dot products grow large in magnitude, pushing the softmax function into regions where it has extremely small gradients [3]. To counteract this effect, we scale the dot products by [math]\displaystyle{ 1/\sqrt{d_k} }[/math].

3.2.2 Multi-Head Attention
Instead of performing a single attention function with [math]\displaystyle{ d_{model} }[/math]-dimensional keys, values and queries, we found it beneficial to linearly project the queries, keys and values [math]\displaystyle{ h }[/math] times with different, learned linear projections to [math]\displaystyle{ d_k }[/math], [math]\displaystyle{ d_k }[/math] and [math]\displaystyle{ d_v }[/math] dimensions, respectively. On each of these projected versions of queries, keys and values we then perform the attention function in parallel, yielding [math]\displaystyle{ d_v }[/math]-dimensional output values. These are concatenated and once again projected, resulting in the final values, as depicted in Figure 2.

 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.

[math]\displaystyle{ MultiHead(Q, K, V) : Concat(head_1, \cdots, head_h)W^O }[/math]

where [math]\displaystyle{ head_i = Attention(QW_i^Q, KW_i^K , VW_i^V) }[/math]

Where the projections are parameter matrices [math]\displaystyle{ W_i^O\in \mathbb{R}^{d_{model}\times d_k} }[/math], [math]\displaystyle{ W_i^K\in \mathbb{R}^{d_{model}\times d_k} }[/math], [math]\displaystyle{ W_i^V\in \mathbb{R}^{d_{model}\times d_k} }[/math] and [math]\displaystyle{ W^O\in \mathbb{R}^{hd_v\times d_{model}} }[/math].

In this words we employ [math]\displaystyle{ h = 8 }[/math] parallel attention layers, or heads. For each of these we use [math]\displaystyle{ dk = dv = d_{model}/h = 64 }[/math]. Due to the reduced dimension of each head, the total computational cost is similar to that of 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:

* 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].

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.

[math]\displaystyle{ FFN=\max(0, xW_1+b_1)W_2+b_2 }[/math] (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. The dimensionality of input and output is [math]\displaystyle{ d_{model}=512 }[/math], and the inner-layer has dimensionality [math]\displaystyle{ d_{ff} = 2048 }[/math].

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 [math]\displaystyle{ d_{model} }[/math]. 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 [math]\displaystyle{ \sqrt{d_{model}} }[/math].

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 [math]\displaystyle{ d_{model} }[/math] as the embeddings, so that the two can be summed. There are many choices of positional encodings, learned and fixed [9].

In this work, we use sine and cosine functions of different frequencies:

[math]\displaystyle{ PE_{(pos,2i)} = \sin(pos/10000^{2i/d_{model}}) }[/math]

[math]\displaystyle{ PE_{(pos,2i+1)} =\cos(pos/10000^{2i/d_{model}}) }[/math]

where [math]\displaystyle{ pos }[/math] is the position and $i$ is the dimension. That is, each dimension of the positional encoding corresponds to a sinusoid. The wavelengths form a geometric progression from [math]\displaystyle{ 2\pi }[/math] to [math]\displaystyle{ 10000 \cdot 2\pi }[/math]. 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 [math]\displaystyle{ k }[/math], [math]\displaystyle{ PE_{pos+k} }[/math] can be represented as a linear function of [math]\displaystyle{ PE_{pos} }[/math].

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.

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 $\left(x_1,\cdots, x_n\right)$ to another sequence of equal length $\left(z1,\cdots, 2\right)$, with $x_i,z_i \in \R^d$, such as a hidden layer in a typical sequence transduction encoder or decoder. Motivating our use of self-attention we consider three desiderata.

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.

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 [11]. Hence we also compare the maximum path length between any two input and output positions in networks composed of the different layer types.

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 $O(n)$ sequential operations. In terms of computational complexity, self-attention layers are faster than recurrent layers when the sequence length $n$ is smaller than the representation dimensionality $d$, which is most often the case with sentence representations used by state-of-the-art models in machine translations, such as word-piece [31] and byte-pair representations. To improve computational performance for tasks involving very long sequences, self-attention could be restricted to considering only a neighborhood of size $r$ in the input sequence centered around the respective output position. This would increase the maximum path length to $O(n/r)$. We plan to investigate this approach further in future work.

A single convolutional layer with kernel width $k < n$ does not connect all pairs of input and output positions. Doing so requires a stack of $O (n/k)$ convolutional layers in the case of contiguous kernels, or $O\left(log, (n)\right)$ in the case of dilated convolutions [15], 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 $k$. Separable convolutions [6], however, decrease the complexity considerably, to $O (k\cdot n\cdot d + n\cdot d^2)$. Even with $k = n$, 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.

As 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.

{|class="wikitable" style="border:1px solid black;border-spacing:1px; text-align:center; margin:1em auto; width:90%" |- ! Layer Type !!Complexity per Layer !!Sequential Operations!! Maximum Path Length |- | Self-Attention || $O(n^2 \cdot d)$ || $O(1)$ || $O(1)$ |- | Recurrent || $O(n \cdot d^2)$|| $O(n)$ || $O(n)$ |- | Convolutional || $O(k\cdot n\cdot d^2)$|| $O(1)$|| $O(log_k (n))$ |- | Self-Attention (restricted)|| $O(r\cdot n\cdot d)$ || $O(1)$|| $O(n/r)$ |- |+ align="bottom" style="caption-side:bottom; text-align:left; font-weight:normal"|Table 1: Maximum path lengths, per-layer complexity and minimum number of sequential operations for different layer types. $n$ is the sequence length, $d$ is the representation dimension, $k$ is the kernel size of convolutions and $r$ the size of the neighborhood in restricted self-attention. |}

5 Training

This section describes the training regime for our models.

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 source-target vocabulary of about 3,7000 tokens. For English-French, we used the significantly larger WMT 2014 English-French dataset consisting of 36M sentences and split tokens into a 32,000 word-piece vocabulary [31]. Sentence pairs were batched together by approximate sequence length. Each training batch contained a set of sentence pairs containing approximately 25,000 source tokens and 25,000 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).

5.3 Optimizer

We used the Adam optimizer 17 with $\beta_1 = 0.9$, $\beta_2 = 0.98$ and $\epsilon = 10^{-9}$. We varied the learning rate over the course of training, according to the formula:

[math]\displaystyle{ lrate=d_{model}^{-0.5}\cdot min\left(step_num^{-0.5}, step_num\cdot warmup_steps^{-1.5}\right) }[/math] (3)

This corresponds to increasing the learning rate linearly for the first warmup_steps training steps, and decreasing it thereafter proportionally to the inverse square root of the step number. We used warmup_steps = 4000.

5.4 Regularization

We employ three types of regularization during training:

Residual Dropout: We apply dropout 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 $P_{drop} = 0.1$.

Label Smoothing: During training, we employed label smoothing of value $\epsilon_{ls} = 0.1$ [3O]. This hurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score.

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.

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 $P_{drop} = 0.1$, instead of $0.3$.

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 a = 0.6 [31]. 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 [31].

 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[4]

{|class="wikitable" style="border:1px solid black;border-spacing:1px; text-align:center; background-color:white; margin:1em auto; width:90%" |- !rowspan="2"|Model !colspan="2"|BLEU !colspan="2"|Training Cost (FLOPs) |- !EN-DE !!EN-FR!! EN-DE !!EN-FR |- | ByteNet [15] ||$23.75$||  ||  ||   |- | Deep-Att + PosUnk [15] ||  ||$39.2$||  || $1.0\cdot 10^{20}$ |- |GNMT + RL [31]||$24.6$||$39.92$|| $2.3\cdot 10^{19}$|| $1.4\cdot 10^{20}$ |- |ConvS2S [8] ||$25.16$|| $40.46$ ||$9.6-10^{18}$|| $1.5\cdot 10^{20}$ |- |style="border-bottom:double"| MoE [26] ||$26.03$ ||$40.56$ ||$2.0\cdot 10^{19}$|| $1.2\cdot 10^{20}$ |- |Deep-Att + PosUnk Ensemble [32]||  || $40.4$||   || $8.0 \cdot 10^{20}$ |- |GNMT + RL Ensemble [31]||$26.30$|| $41.16$ ||$1.8\cdot 10^{20}$|| $1.1\cdot 10^{21}$ |- |ConvS2S Ensemble [8] ||$26.36$ ||$\mathbf{41.29}$|| $7.7\cdot 10^{19}$|| $1.2\cdot 10^{21}$ |- |style="border-bottom:double"|Transformer (base model) ||$27.3$|| $38.1$ |colspan="2"|$\mathbf{3.3 \cdot 10^{18}}$ |- |Transformer (big) ||$\mathbf{28.4}$ ||$\mathbf{41.0}$ |colspan="2"|$2.3\cdot 10^{19}$ |- |+ align="bottom" style="caption-side:bottom; text-align:left; font-weight:normal"|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. |}

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.

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 While single-head attention is 0.9 BLEU worse than the best setting, quality also drops off with too many heads.

In Table 3 rows (B), we observe that reducing the attention key size $d_k$ 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 (C) 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 8, and observe nearly identical results to the base model.

{|class="wikitable" style="border:1px solid black;border-spacing:1px; text-align:center; background-color:white; margin:1em auto; width:90%" |- ! !!$N$ !!$d_{model}$!! $d_{ff}$ !! h!! $d_k$!!$d_v$!! $P_{drop}$!! $\epsilon_{ls}$!!train steps !! PPL

(dev)!!BLEU

(dev) !! params$\times 10^6$ |- ! base | 6 ||512 ||2,048||8 ||64 ||64 ||0.1 ||O.1|| 100K || 4.92 ||25.8|| 65 |- !rowspan="4"| (A) | || ||  ||1 ||512 ||512|| || || ||5.29|| 24.9||  |- | || ||  ||4||128||128|| || || ||5.00 ||25.5||  |- | || ||  ||16||32||32|| || || ||4.91|| 25.8||  |- | || ||  ||32||16||16|| || ||  ||5.01 ||25.4||  |- !rowspan="3"| (B) |- | || ||  || || 16|| || ||  || ||5.16 ||25.1|| 58 |- | || ||  || ||32|| || ||  || || 5.01|| 25.4 ||60 |- !rowspan="7"|(C) |2|| || || || || || || || || 6.11 ||23.7|| 36 |- |4 || || || || || || || || ||5.19|| 25.3|| 50 |- |8 || || || || || || || || ||4.88||25.5|| 80 |- | ||256|| || ||32||32|| || || ||5.75|| 24.5|| 28 |- | ||1024|| || ||128||128|| || || ||4.66|| 26.0 ||168 |- | || ||1024|| || || || || || ||5.12|| 25.4|| 53 |- | || ||4096|| || || || || || || 4.75|| 26.2|| 90 |- !rowspan="4"|(D) | || || || || || ||0.0|| || || 5.77||24.6||  |- | || || || || || ||0.2|| || || 4.95|| 25.5||  |- | || || || || || || ||0.0|| ||4.67 ||25.3||  |- | || || || || || || ||0.2|| ||5.47|| 25.7||  |- !(E) |colspan="9"|positional embedding instead of sinusoids|| 4.92|| 25.7||  |- ! big | 6 ||1024|| 4096|| 16|| || || 0.3|| || 300K || 4.33|| 26.4 ||213 |- |+ align="bottom" style="caption-side:bottom; text-align:left; font-weight:normal"|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. |}

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.

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.

 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.

The code we used to train and evaluate our models is available at https://github.com/tensorflow/tensor2tensor

Acknowledgements

We are grateful to Nal Kalchbrenner and Stephan Gouws for their fruitful comments, corrections and inspiration.

Footnotes

  1. ↑ https://github.com/tensorflow/tensor2tensor
  2. ↑ http://nlp.seas.harvard.edu/2018/04/03/attention.html
  3. ↑ To illustrate why the dot products get large, assume that the components of [math]\displaystyle{ q }[/math] and [math]\displaystyle{ k }[/math] are independent random variables with mean 0 and variance 1. Then their dot product, [math]\displaystyle{ q\cdot k=\displaystyle \sum^{d_k}_{i=1}q_ik_i }[/math], has mean 0 and variance [math]\displaystyle{ d_k }[/math].
  4. ↑ We used values of 2.8, 3.7, 6.0 and 9.5 TFLOPS for K80, K40, M40 and P100, respectively.

References

BibTeX

@inproceedings{2017_AttentionisallYouNeed,
  author    = {Ashish Vaswani and
 [[Noam Shazeer]] and
               Niki Parmar and
               Jakob Uszkoreit and
               Llion Jones and
               Aidan N. Gomez and
               Lukasz Kaiser and
               Illia Polosukhin},
  editor    = {Isabelle Guyon and
               Ulrike von Luxburg and
               Samy Bengio and
               Hanna M. Wallach and
               Rob Fergus and
               S. V. N. Vishwanathan and
               Roman Garnett},
  title     = {Attention is All you Need},
  booktitle = {Advances in Neural Information Processing Systems 30: Annual Conference
               on Neural Information Processing Systems 2017, 4-9 December 2017,
               Long Beach, CA, USA},
  pages     = {5998--6008},
  year      = {2017},
  url       = {http://papers.nips.cc/paper/7181-attention-is-all-you-need},
}

 AuthorvolumeDate ValuetitletypejournaltitleUrldoinoteyear
2017 AttentionisallYouNeedLukasz Kaiser
Noam Shazeer
Ashish Vaswani
Niki Parmar
Jakob Uszkoreit
Llion Jones
Aidan N Gomez
Illia Polosukhin		Attention is all You Need2017
Retrieved from "http://www.gabormelli.com/RKB/index.php?title=2017_AttentionisallYouNeed&oldid=903001"