Attention Is All You Need
Abstract The dominant sequence transduction models are based on complex recurrent or convolutional n
eural networks that include an encoder and a decoder. The best performing models also connect the en
coder 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 entir
ely. 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 BL
EU on the WMT 2014 Englishto-German translation task, improving over the existing best results, incl
uding ensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model estab
lishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight G
PUs, 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 par
sing both with large and limited training data. 1 Introduction Recurrent neural networks, long short
-term memory and gated recurrent neural networks in particular, have been firmly established as st
ate of the art approaches in sequence modeling and Equal contribution. Listing order is random. Jak
ob 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 i
n every aspect of this work. Noam proposed scaled dot-product attention, multi-head attention and th
e 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 initia
l codebase, and efficient inference and visualizations. Lukasz and Aidan spent countless long days d
esigning various parts of and implementing tensor2tensor, replacing our earlier codebase, greatly im
proving results and massively accelerating our research. Work performed while at Google Brain. Wor
k performed while at Google Research. 31st Conference on Neural Information Processing Systems (NIPS
2017), Long Beach, CA, USA.transduction problems such as language modeling and machine translation
[35, 2, 5]. Numerous efforts have since continued to push the boundaries of recurrent language model
s and encoder-decoder architectures. Recurrent models typically factor computation along the symbol
positions of the input and output sequences. Aligning the positions to steps in computation time, th
ey 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 ex
amples, which becomes critical at longer sequence lengths, as memory constraints limit batching acro
ss examples. Recent work has achieved significant improvements in computational efficiency through f
actorization tricks and conditional computation , while also improving model performance in case of
the latter. The fundamental constraint of sequential computation, however, remains. Attention mechan
isms 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 se
quences [2, 19]. In all but a few cases , however, such attention mechanisms are used in conjunction
with a recurrent network. In this work we propose the Transformer, a model architecture eschewing r
ecurrence 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.
Background The goal of reducing sequential computation also forms the foundation of the Extended Neu
ral GPU , ByteNet and ConvS2S , all of which use convolutional neural networks as basic building bl
ock, computing hidden representations in parallel for all input and output positions. In these model
s, 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 . 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 desc
ribed in section 3.2. Self-attention, sometimes called intra-attention is an attention mechanism rel
ating 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, ab
stractive summarization, textual entailment and learning task-independent sentence representations.
End-to-end memory networks are based on a recurrent attention mechanism instead of sequencealigned r
ecurrence and have been shown to perform well on simple-language question answering and language mod
eling tasks . 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, m
otivate self-attention and discuss its advantages over models such as and.
Model Architecture Most competitive neural sequence transduction models have an encoder-decoder stru
cture. Here, the encoder maps an input sequence of symbol representations to a sequence of continuou
s representations. Given z, the decoder then generates an output sequence of symbols o
ne element at a time. At each step the model is auto-regressive , consuming the previously generated
symbols as additional input when generating the next. The Transformer follows this overall architec
ture using stacked self-attention and point-wise, fully connected layers for both the encoder and de
coder, shown in the left and right halves of Figure 1, respectively. 2 Figure 1: The Transformer - m
odel architecture.
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 c
onnected feed-forward network. We employ a residual connection around each of the two sub-layers, f
ollowed by layer normalization . That is, the output of each sub-layer is LayerNorm(x + Sublayer(x))
, where Sublayer(x) is the function implemented by the sub-layer itself. To facilitate these residua
l connections, all sub-layers in the model, as well as the embedding layers, produce outputs of dime
nsion dmodel = 512.
Decoder: The decoder is also composed of a stack of N = 6 identical layers. In addition to the two s
ub-layers in each encoder layer, the decoder inserts a third sub-layer, which performs multi-head at
tention 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 s
ub-layer in the decoder stack to prevent positions from attending to subsequent positions. This mask
ing, combined with fact that the output embeddings are offset by one position, ensures that the pred
ictions for position i can depend only on the known outputs at positions less than 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 t
he query with the corresponding key. 3 Scaled Dot-Product Attention Multi-Head Attention Figure 2: (
left) Scaled Dot-Product Attention. (right) Multi-Head Attention consists of several attention layer
s running in parallel. 3.2.1 Scaled Dot-Product Attention We call our particular attention "Scaled D
ot-Product Attention" (Figure 2). The input consists of queries and keys of dimension dk, and values
of dimension dv. We compute the dot products of the query with all keys, divide each by dk, and ap
ply a softmax function to obtain the weights on the values. In practice, we compute the attention fu
nction on a set of queries simultaneously, packed together into a matrix Q. The keys and values are
also packed together into matrices K and V . We compute the matrix of outputs as: Attention(Q, K, V
) = softmax(QKT dk )V (1) The two most commonly used attention functions are additive attention , an
d dot-product (multiplicative) attention. Dot-product attention is identical to our algorithm, excep
t for the scaling factor of 1 dk . Additive attention computes the compatibility function using a f
eed-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 implemen
ted using highly optimized matrix multiplication code. While for small values of dk the two mechanis
ms perform similarly, additive attention outperforms dot product attention without scaling for large
r values of dk . We suspect that for large values of dk, the dot products grow large in magnitude, p
ushing the softmax function into regions where it has extremely small gradients 4 . To counteract th
is effect, we scale the dot products by 1 dk . 3.2.2 Multi-Head Attention Instead of performing a si
ngle attention function with dmodel-dimensional keys, values and queries, we found it beneficial to
linearly project the queries, keys and values h times with different, learned linear projections to
dk, dk and dv dimensions, respectively. On each of these projected versions of queries, keys and val
ues we then perform the attention function in parallel, yielding dv-dimensional output values. These
are concatenated and once again projected, resulting in the final values, as depicted in Figure 2.
To illustrate why the dot products get large, assume that the components of q and k are independent
random variables with mean 0 and variance 1. Then their dot product. In this work we employ h = 8 pa
rallel attention layers, or heads. For each of these we use dk = dv = dmodel/h = 64. Due to the redu
ced 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 diff
erent 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 t
he decoder to attend over all positions in the input sequence. This mimics the typical encoder-decod
er attention mechanisms in sequence-to-sequence models such as. The encoder contains self-attention
layers. In a self-attention layer all of the keys, values and queries come from the same place, in t
his case, the output of the previous layer in the encoder. Each position in the encoder can attend t
o all positions in the previous layer of the encoder. • Similarly, self-attention layers in the deco
der 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-reg
ressive property. We implement this inside of scaled dot-product attention by masking out (setting t
o −∞) all values in the input of the softmax which correspond to illegal connections. See Figure 2.
3.3 Position-wise Feed-Forward Networks In addition to attention sub-layers, each of the layers in o
ur encoder and decoder contains a fully connected feed-forward network, which is applied to each pos
ition separately and identically. This consists of two linear transformations with a ReLU activation
in between. While the linear transformations are the same across different positions, they use diff
erent parameters from layer to layer. Another way of describing this is as two convolutions with ker
nel size 1. The dimensionality of input and output is dmodel = 512, and the inner-layer has dimensio
nality df f = 2048.
3.4 Embeddings and Softmax
Similarly to other sequence transduction models, we use learned embeddings to convert the input toke
ns and output tokens to vectors of dimension dmodel. We also use the usual learned linear transforma
tion and softmax function to convert the decoder output to predicted next-token probabilities. In ou
r model, we share the same weight matrix between the two embedding layers and the pre-softmax linear
transformation. In the embedding layers, we multiply those weights by dmodel.
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 t
he 5 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 ke
rnel size of convolutions and r the size of the neighborhood in restricted self-attention. Layer Typ
e Complexity per Layer tokens in the sequence. To this end, we add "positional encodings" to the inp
ut embeddings at the bottoms of the encoder and decoder stacks. The positional encodings have the sa
me dimension dmodel as the embeddings, so that the two can be summed. There are many choices of posi
tional encodings, learned and fixed . In this work, we use sine and cosine functions of different fr
equencies: where pos is the position and i is the dimension. That is, each dimension of the position
al encoding corresponds to a sinusoid.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, P Epos+k ca
n be represented as a linear function of P Epos. We also experimented with using learned positional
embeddings 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 convolution
al layers commonly used for mapping one variable-length sequence of symbol representations (x1, ...,
xn) to another sequence of equal length (z1, ..., zn), 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 t
he 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 su
ch dependencies is the length of the paths forward and backward signals have to traverse in the netw
ork. The shorter these paths between any combination of positions in the input and output sequences,
the easier it is to learn long-range dependencies . Hence we also compare the maximum path length b
etween any two input and output positions in networks composed of the different layer types. As note
d in Table 1, a self-attention layer connects all positions with a constant number of sequentially e
xecuted operations, whereas a recurrent layer requires O(n) sequential operations. In terms of compu
tational 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 r
epresentations used by state-of-the-art models in machine translations, such as word-piece 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 6 the input seq
uence 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 w
ith 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(logk(n)) in the case o
f dilated convolutions , 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 , however, decrease the complexity considerably. Even with k = n, however, t
he 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 c
ould 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 perfo
rm different tasks, many appear to exhibit behavior related to the syntactic and semantic structure
of the sentences.