본문 바로가기
딥러닝/논문리뷰

attention is all you need 한글 번역

by 달죽 2020. 11. 2.
반응형



Abstract

 

The dominant sequence transduction models are based on complex recurrent orconvolutional neural networks that include an encoder and a decoder
지배적인 시퀀스 변환 모델은 인코더 및 디코더를 포함하는 복잡한 반복 또는 컨볼루션 신경망을 기반으로 합니다


The bestperforming models also connect the encoder and decoder through an attentionmechanism
또한 가장 우수한 성능의 모델은 주의 메커니즘을 통해 인코더와 디코더를 연결합니다


" We propose a new simple network architecture, the Transformer,based solely on attention mechanisms, dispensing with recurrence and convolutions entirely"

"우리는 단지 주의 메커니즘에 기반을 두고, recurrence 와 컨볼루션으로 분배하는 새로운 단순한 네트워크 아키텍처인 트랜스포머를 제안한다."


Experiments on two machine translation tasks show these models tobe superior in quality while being more parallelizable and requiring signicantly less time to train
두 가지 기계 번역 작업에 대한 실험은 이러한 모델들이 더 병렬화 가능하고 훈련하는데 시간이 더 적게 걸리면서

품질이 우수하다는 것을 보여준다.


Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles, by over 2 BLEU"
우리의 모델은 28.4 을 달성한다 WMT 2014 영어-독일 번역 작업에 대한 2 BLEU 이상 ensembles 등 기존 최상의 결과보다 개선되었다. 


On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.0 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature"
WMT 2014 영어 대 프랑스어 번역 작업에서 우리 모델은 41점의 새로운 단일 모델 BLEU 8개의 GPU로 3.5일 걸렸다, 이것은 문헌에서 가장 좋은 모델의 훈련 비용의 극히 일부다. "

Introduction

 

Recurrent neural networks, long short-term memory [12] and gated recurrent [7] neural networksin particular, have been rmly established as state of the art approaches in sequence modeling andtransduction problems such as language modeling and machine translation [29, 2, 5]"
"특히, 언어 모델링 및 기계 번역과 같은 시퀀스 모델링 및 변환 문제에서 기술적인 접근의 상태로서, 장기간의 기억[12]과 게이트된 재귀성[7] 신경망이 확고히 확립되었다[29, 2, 5]"


" Numerousefforts have since continued to push the boundaries of recurrent language models and encoder-decoderarchitectures [31, 21, 13]"
"그 후 수많은 사람들이 반복적인 언어 모델과 인코더 디코더 구조의 경계를 계속 밀어붙였다[31, 21, 13]"
Equal contribution
동등한 기여


Listing order is random
나열 순서는 무작위입니다


Jakob proposed replacing RNNs with self-attention and startedthe effort to evaluate this idea
야콥은 RNN을 자기주의로 대체할 것을 제안했고 이 아이디어를 평가하기 위한 노력을 시작했다.
" Ashish, with Illia, designed and implemented the rst Transformer models andhas been crucially involved in every aspect of this work"
일리아와 함께 애쉬쉬는 최초의 트랜스포머 모델을 설계하고 구현했으며 이 작업의 모든 측면에 결정적으로 관여했다.


" 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"
"노암은 점-제품 주의력, 다중 머리 집중력, 파라미터 없는 위치 표현을 확장하여 거의 모든 세부 사항에 관여하는 다른 사람이 되었다."


" Niki designed, implemented, tuned and evaluated countless model variants in our original codebase andtensor2tensor"
"니키는 원래 코드베이스와 텐소르2텐서에서 수많은 모델 변형을 설계, 구현, 튜닝 및 평가했다."


" Llion also experimented with novel model variants, was responsible for our initial codebase, andefcient inference and visualizations"
"Llion은 또한 새로운 모델 변형을 실험했고, 우리의 초기 코드베이스, 효율적인 추론 및 시각화를 담당했다."


" Lukasz and Aidan spent countless long days designing various parts of andimplementing tensor2tensor, replacing our earlier codebase, greatly improving results and massively acceleratingour research"
루카스와 에이단은 수많은 긴 하루를 텐서2텐서의 다양한 부분을 설계하고 구현하며 우리의 초기 코드베이스를 대체하여 결과를 크게 개선하고 대규모로 연구를 가속화했다.


†Work performed while at Google Brain
구글 브레인에서 수행된 작업
‡Work performed while at Google Research
구글 리서치에서 수행된 작업


"31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA"
미국 캘리포니아 주 롱비치 신경정보처리시스템(NIPS 2017) 31차 컨퍼런스


Recurrent models typically factor computation along the symbol positions of the input and outputsequences
반복 모델은 일반적으로 입력 및 출력 시퀀스의 기호 위치를 따라 계산을 계수화합니다


" Aligning the positions to steps in computation time, they generate a sequence of hiddenstates ht, as a function of the previous hidden state ht−1 and the input for position t"
"위치들을 계산 시간의 단계들에 정렬하여, 그들은 이전의 숨겨진 상태 ht1 및 위치 t에 대한 입력의 함수로서 숨겨진 상태 ht의 시퀀스를 생성한다."


" This inherentlysequential nature precludes parallelization within training examples, which becomes critical at longersequence lengths, as memory constraints limit batching across examples"
이러한 본질적으로 순차적인 성질은 메모리 제약으로 인해 예제에 대한 배치가 제한되기 때문에 더 긴 시퀀스 길이에서 중요한 훈련 예제 내에서 병렬화를 배제한다.


" Recent work has achievedsignicant improvements in computational efciency through factorization tricks [18] and conditionalcomputation [26], while also improving model performance in case of the latter"
최근 연구는 인수 분해 트릭 [18]과 조건부 계산 [26]을 통해 계산 효율을 크게 향상 시켰으며 후자의 경우 모델 성능을 향상 시켰습니다.


" The fundamentalconstraint of sequential computation, however, remains"

그러나 순차적 계산의 기본 제약은 여전히 남아 있다.


"Attention mechanisms have become an integral part of compelling sequence modeling and transduc-tion models in various tasks, allowing modeling of dependencies without regard to their distance inthe input or output sequences [2, 16]"
"주의 메커니즘은 다양한 작업에서 설득력 있는 시퀀스 모델링 및 변환-동작 모델의 필수적인 부분이 되어 입력 또는 출력 시퀀스[2, 16]에서의 거리에 관계없이 의존성의 모델링을 가능하게 한다."


" In all but a few cases [22], however, such attention mechanismsare used in conjunction with a recurrent network"
"그러나 몇 가지 경우를 제외하고, 이러한 주의 메커니즘은 반복적인 네트워크와 함께 사용된다."


"In this work we propose the Transformer, a model architecture eschewing recurrence and insteadrelying entirely on an attention mechanism to draw global dependencies between input and output"본 논문에서는 변환기를 제안한다. 변환기는 재귀를 회피하고 대신 입출력 사이의 글로벌 의존성을 도출하기 위한 주의 메커니즘에 전적으로 의존하는 모델 아키텍처이다.


The Transformer allows for signicantly more parallelization and can reach a new state of the art intranslation quality after being trained for as little as twelve hours on eight P100 GPUs
트랜스포머는 훨씬 더 병렬화를 허용하며 8개의 P100 GPU에서 12시간 동안 훈련된 후 새로운 예술 번역 품질에 도달할 수 있습니다.


Background

 

The goal of reducing sequential computation also forms the foundation of the Extended Neural GPU[20], ByteNet [15] and ConvS2S [8], all of which use convolutional neural networks as basic buildingblock, computing hidden representations in parallel for all input and output positions"
"2 순차 연산을 줄이는 목표는 또한 확장된 신경 GPU [20], ByteNet [15] ConvS2S [8]의 기초를 형성하며, 모두 컨볼루션 신경망을 기본 구성 요소로 사용하고 모든 입력 및 출력 위치에 대해 숨겨진 표현을 병렬로 계산합니다."


" In these models,the number of operations required to relate signals from two arbitrary input or output positions growsin the distance between positions, linearly for ConvS2S and logarithmically for ByteNet"
"이러한 모델에서 두 임의의 입력 또는 출력 위치의 신호를 연관시키는 데 필요한 연산 수는 ConvS2S에 대해 선형으로, ByteNet에 대해 로그로 증가한다."


This makesit more difcult to learn dependencies between distant positions [11]
이것은 먼 위치 사이의 의존성을 배우는 것을 더 어렵게 만듭니다 [11]


" In the Transformer this isreduced to a constant number of operations, albeit at the cost of reduced effective resolution dueto averaging attention-weighted positions, an effect we counteract with Multi-Head Attention asdescribed in section 3.2"
"트랜스포머에서 이것은 비록 평균적인 주의 가중치 위치로 인한 효과적인 해상도 감소의 비용에도 불구하고 일정한 수의 작업으로 감소되며, 이는 제3.2 조에 기술된 다중 헤드 주의와 상쇄되는 효과다."

"Self-attention, sometimes called intra-attention is an attention mechanism relating different positionsof a single sequence in order to compute a representation of the sequence"
때로는 intra- 주의력 라고도하는 자기 주의력은 시퀀스의 표현을 계산하기 위해 단일 시퀀스의 다른 위치를 관련시키는 주의 메커니즘입니다.


" Self-attention has beenused successfully in a variety of tasks including reading comprehension, abstractive summarization,textual entailment and learning task-independent sentence representations [4, 22, 23, 19]"
"자기주의는 읽기 이해, 추상적 요약, 텍스트 수반 및 학습 과제 독립 문장 표현을 포함한 다양한 과제에서 성공적으로 사용되었습니다 [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 andlanguage modeling tasks [28]
엔드 투 엔드 메모리 네트워크는 시퀀스 정렬된 반복 대신 반복적인 주의 메커니즘에 기반을 두고 있으며 단순 언어 질문 응답 및 언어 모델링 작업에서 잘 수행되는 것으로 나타났습니다 [28]


"To the best of our knowledge, however, the Transformer is the rst transduction model relyingentirely on self-attention to compute representations of its input and output without using sequence-aligned RNNs or convolution"
"그러나 우리가 아는 한, 변환기는 시퀀스 정렬된 RNN이나 컨볼루션을 사용하지 않고도 입력과 출력의 표현을 계산하기 위해 자기 주의에 의존하는 최초의 변환 모델이다."


" In the following sections, we will describe the Transformer, motivateself-attention and discuss its advantages over models such as [14, 15] and [8]"
"다음 절에서는 트랜스포머를 설명하고, 자기주의를 자극하며, [14, 15] [8]과 같은 모델에 비해 그 장점에 대해 논의할 것이다."

 


Model Architecture

 

Most competitive neural sequence transduction models have an encoder-decoder structure [5, 2, 29]"
"3 모델 아키텍처의 가장 경쟁적인 신경 시퀀스 변환 모델은 인코더 디코더 구조 [5, 2, 29]를 갖는다."


"Here, the encoder maps an input sequence of symbol representations (x1, ... , xn) to a sequenceof continuous representations z = (z1, ... zn)"
"여기서, 인코더는 심볼 표현(x1, ... , xn) 연속 표현 z = (z1 ,...,  zn)"

" Given z, the decoder then generates an outputsequence (y1, ... , ym) of symbols one element at a time"
"주어진 z, 디코더는 출력 시퀀스(y1, ... , ym)의 심볼은 한 번에 한 요소씩"

" At each step the model is auto-regressive[9], consuming the previously generated symbols as additional input when generating the next"
"각 단계에서 모델은 자동 회귀성 [9]이며, 다음 생성시 이전에 생성된 기호를 추가 입력으로 소비합니다."


"The Transformer follows this overall architecture using stacked self-attention and point-wise, fullyconnected layers for both the encoder and decoder, shown in the left and right halves of Figure 1,respectively"


변환기는 그림 1의 왼쪽과 오른쪽 반쪽에 각각 표시된 인코더와 디코더 모두에 대해 스택된 자체 주의력과 점별로 완전히 연결된 레이어를 사용하여 전체 아키텍처를 따릅니다.

3. 1 Encoder and Decoder Stacks

Encoder:

The encoder is composed of a stack of N = 6 identical layers
1 엔코더 및 디코더 스택스엔코더: 엔코더는 N = 6개의 동일한 레이어 스택으로 구성된다.


Each layer has twosub-layers
각 층은 2개의 서브층을 갖는다


" The rst is a multi-head self-attention mechanism, and the second is a simple, position-2 Figure 1: The Transformer - model architecture"
첫 번째는 다중 헤드 자기 주의 메커니즘이고 두 번째는 단순하고 위치 2 그림 1 : 변환기 - 모델 아키텍처입니다.


wise fully connected feed-forward network
현명한 완전 연결 피드-포워드 네트워크


" We employ a residual connection [10] around each ofthe two sub-layers, followed by layer normalization [1]"
"우리는 두 하위 계층 각각에 대한 잔여 연결 [10]을 사용하고, 그 다음에 계층 정규화 [1]을 사용한다."


" That is, the output of each sub-layer isLayerNorm(x + Sublayer(x)), where Sublayer(x) is the function implemented by the sub-layeritself"
", 각 하위 계층의 출력은 LayerNorm(x + Sublayer(x)이며, 여기서 하위 계층(x)은 하위 계층 자체에 의해 구현되는 기능이다."


" To facilitate these residual connections, all sub-layers in the model, as well as the embeddinglayers, produce outputs of dimension dmodel = 512"
이러한 잔차 연결을 용이하게 하기 위해 모델의 모든 하위 계층과 임베딩 계층은 차원 dmodel = 512의 출력을 생성합니다


Decoder:

 

The decoder is also composed of a stack of N = 6 identical layers
디코더: 디코더는 또한 N = 6개의 동일한 층의 스택으로 구성된다.
" In addition to the twosub-layers in each encoder layer, the decoder inserts a third sub-layer, which performs multi-headattention over the output of the encoder stack"
디코더는 각 인코더 계층의 2개의 서브 계층 외에도 인코더 스택의 출력에 대해 다중 헤드 어텐션을 수행하는 제 3 서브 계층을 삽입한다
" Similar to the encoder, we employ residual connectionsaround each of the sub-layers, followed by layer normalization"
"인코더와 유사하게, 우리는 각 하위 계층에 대해 잔여 연결을 사용하고, 그 다음에 계층 정규화를 사용한다."
We also modify the self-attentionsub-layer in the decoder stack to prevent positions from attending to subsequent positions
또한 디코더 스택의 자기 주의 서브 레이어를 수정하여 위치가 후속 위치에 참석하지 않도록 합니다
" Thismasking, combined with fact that the output embeddings are offset by one position, ensures that thepredictions for position i can depend only on the known outputs at positions less than i"
이 마스크는 출력 임베딩이 하나의 위치로 오프셋된다는 사실과 결합하여 위치 예측이 i보다 작은 위치에서 알려진 출력에만 의존할 수 있음을 보장한다.

 

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"
"주의 기능은 쿼리 및 키 값 쌍 집합을 출력에 매핑하는 것으로 설명될 수 있으며, 여기서 쿼리, , 값 및 출력은 모두 벡터이다."
"The output is computed as a weighted sumof the values, where the weight assigned to each value is computed by a compatibility function of thequery with the corresponding key"
"출력은 값들의 가중 합으로 계산되며, 여기서 각 값에 할당된 가중치는 해당 키와의 질의의 호환성 함수에 의해 계산된다."

 

3.2.1 Scaled Dot-Product Attention

 

 

We call our particular attention ""Scaled Dot-Product Attention"" (Figure 2)"
"1 스케일 도트-프로덕트 주의 우리는 우리의 특별한 관심을 ""스케일 도트-프로덕트 주의""라고 부른다(그림 2)"


" The input consists ofqueries and keys of dimension dk, and values of dimension dv"
입력은 차원 dk의 질의 및 키와 차원 dv의 값으로 구성됩니다


We compute the dot products of the3 Scaled Dot-Product AttentionMulti-Head AttentionFigure 2: (left) Scaled Dot-Product Attention
3Scaled Dot-Product AttentionMulti-Head AttentionFigure 2: (왼쪽) 스케일 도트-프로덕트 주의의 도트 제품을 계산한다.
(right) Multi-Head Attention consists of severalattention layers running in parallel
(오른쪽) 다중 헤드 주의는 병렬로 실행되는 여러 주의 계층으로 구성된다.


"√query with all keys, divide each by dk, and apply a softmax function to obtain the weights on the values.
"모든 키로 질의, 각 바이값을 나누기"


"In practice, we compute the attention function on a set of queries simultaneously, packed togetherinto a matrix Q"
"실제로, 우리는 매트릭스 Q로 함께 포장된 일련의 쿼리에 대한 주의 함수를 동시에 계산한다."


The keys and values are also packed together into matrices K and V
키와 값은 또한 행렬 KV로 함께 포장된다.


" We computethe matrix of outputs as:

 

Attention(Q, K, V ) = softmax(QK^T/ √dk)V (1)

 

 

The two most commonly used attention functions are additive attention [2], and dot-product (multi-plicative) attention"
"출력 행렬을 다음과 같이 계산하고 소프트맥스 함수를 적용하여 Attention(Q, K, V) = softmax(QK Tdk)V 에서 가중치를 구한다. 가장 많이 사용되는 두 가지 주의 기능은 첨가 주의력[2], 그리고 점 생성물(다중첨가) 주의력이다."

 

" Dot-product attention is identical to our algorithm, except for the scaling factor of  1/√dk "
-제품 주의력은 스케일링 인자를 제외하고 우리의 알고리즘과 동일하다.


Additive attention computes the compatibility function using a feed-forward network witha single hidden layer
추가 주의는 단일 숨겨진 계층을 가진 피드 포워드 네트워크를 사용하여 호환성 함수를 계산합니다


" While the two are similar in theoretical complexity, dot-product attention ismuch faster and more space-efcient in practice, since it can be implemented using highly optimizedmatrix multiplication code"
"두 가지가 이론적 복잡성에서는 유사하지만, -제품 주의는 매우 빠르고 공간 효율적이다. 왜냐하면 그것은 고도로 최적화된 행렬 곱셈 코드를 사용하여 구현될 수 있기 때문이다."


"While for small values of dk the two mechanisms perform similarly, additive attention outperforms dot product attention without scaling for larger values of dk [3]"
"dk의 작은 값에 대해 두 메커니즘은 유사하게 수행되지만, 첨가 주의는 dk [3]의 더 큰 값을 위해 스케일링하지 않고 제품 주의력을 능가한다."


" We suspect that for large values ofdk, the dot products grow large in magnitude, pushing the softmax function into regions where it hasextremely small gradients (4)"
"우리는 dk의 큰 값의 경우, 점 생성물이 크기가 커져서 소프트맥스 기능을 극히 작은 기울기 4가 있는 지역으로 밀어넣는 것으로 의심한다."

 

" To counteract this effect, we scale the dot products by 1/√dk"
이러한 효과를 상쇄하기 위해 우리는 점 생성물을 1/√dk로 확장한다

 

 

 

(4)

그리고 점 생성물인 q · k = ∑ qi x ki 왜 점 생성물이 커지는지를 설명하기 위해 q와 k의 성분이 독립적인 랜덤 = 1 qiki이고 평균 0과 분산 dk를 갖는다고 가정합니다.

 


3.2.2  Multi-Head Attention

 

Instead of performing a single attention function with dmodel-dimensional keys, values and queries,we found it benecial to linearly project the queries, keys and values h times with different, learnedlinear projections to dk, dk and dv dimensions, respectively"


"2 다차원 주의력 dmodel-dimensional key, values, queries로 단일 주의 기능을 수행하는 대신, 우리는 dk, dk, dv 치수에 각각 다른 학습된 선형 투영으로 쿼리, , h배를 선형적으로 투영하는 것이 유익하다는 것을 발견했다."


" On each of these projected versions ofqueries, keys and values we then perform the attention function in parallel, yielding dv-dimensionaloutput values"
"이러한 각 투영된 버전의 질의, 키 및 값에서 우리는 dv 차원 출력 값을 산출하는 병렬로 주의 함수를 수행합니다"


" These are concatenated and once again projected, resulting in the nal values, asdepicted in Figure 2"
이것들은 연결되고 다시 한번 투영되어 그림 2에서 묘사된 최종 값을 산출한다.


Multi-head attention allows the model to jointly attend to information from different representationsubspaces at different positions
다중 머리 주의는 모델이 서로 다른 위치에서 서로 다른 표현 서브 스페이스의 정보를 공동으로 처리할 수 있게 한다.


" With a single attention head, averaging inhibits this"
한 개의 주의력으로 평균화하면 이 현상이 억제됩니다



 

 

주의 (QW Qi, KW Ki, V W Vi) 투영이 매개 변수 행렬  W Qand W O Rhdvdmodel 인 경우 "

 

 

이 작업에서 우리는  h = 8 개의 평행 주의 계층 또는 헤드를 사용한다 .각각의 경우 우리는  dv = dmodel/h = 64 를 사용했다 .각 헤드의 치수가 감소했기 때문에 전체 계산 비용은 전체 치수를 가진 단일 머리 주의와 유사합니다


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"
"3 모델 주의의 응용 변환기는 세 가지 다른 방법으로 다중 헤드 주의력을 사용합니다.

 

""인코더 디코더 주의""레이어에서 쿼리는 이전 디코더 레이어에서 나오고 메모리 키와 값은 인코더의 출력에서 나옵니다."
This allows everyposition in the decoder to attend over all positions in the input sequence
이를 통해 디코더의 모든 위치가 입력 시퀀스의 모든 위치에 참석할 수 있습니다
" This mimics thetypical encoder-decoder attention mechanisms in sequence-to-sequence models such as[31, 2, 8]"
"이것은 [31, 2, 8]과 같은 시퀀스 대 시퀀스 모델에서의 전형적인 인코더 디코더 주의 메커니즘을 모방합니다."


 The encoder contains self-attention layers
● 인코더에는 자기 주의 계층이 포함되어 있다.
" In a self-attention layer all of the keys, valuesand queries come from the same place, in this case, the output of the previous layer in theencoder"
"자기 주의 계층에서 모든 키, 값 및 쿼리는 동일한 장소에서 나옵니다.이 경우 인코딩기에서 이전 계층의 출력이 나타납니다"
Each position in the encoder can attend to all positions in the previous layer of theencoder
인코더의 각 위치는 인코더의 이전 계층의 모든 위치에 참석할 수 있습니다


Similarly, self-attention layers in the decoder allow each position in the decoder to attend toall positions in the decoder up to and including that position"
마찬가지로 디코더의 자기 주의 계층은 디코더의 각 위치가 디코더의 모든 위치에 해당 위치를 포함하도록 허용합니다.

 

We need to prevent leftwardinformation ow in the decoder to preserve the auto-regressive property
자동 회귀 특성을 보존하기 위해 디코더에서의 좌향 정보 흐름을 방지해야 합니다
We implement thisinside of scaled dot-product attention by masking out (setting to −∞) all values in the inputof the softmax which correspond to illegal connections
우리는 불법적인 연결에 해당하는 소프트맥스의 입력에 있는 모든 값을 마스킹하여(설정 ) 스케일된 도트-제품 주의의 내부에서 이를 구현한다.

 

See Figure 2
그림 2를 참조하십시오

 


3.3 Position-wise Feed-Forward Networks

 

In addition to attention sub-layers, each of the layers in our encoder and decoder contains a fullyconnected feed-forward network, which is applied to each position separately and identically"
"3 위치별 피드-포워드 네트워크 주의 하위 계층 외에, 우리의 인코더 및 디코더의 각 계층에는 완전히 연결된 피드-포워드 네트워크가 포함되어 있으며, 이 네트워크는 각 위치에 개별적으로 동일하게 적용된다."


Thisconsists of two linear transformations with a ReLU activation in between
이 두 선형 변환의 이점과 그 사이에 RELU 활성화

 

 

While the linear transformations are the same across different positions, they use different parametersfrom layer to layer"
"FFN(x) = max(0, xW1 + b1)W2 + b2 (2) 선형 변환은 서로 다른 위치에서 동일하지만, 층마다 다른 매개 변수를 사용한다."

 


Another way of describing this is as two convolutions with kernel size 1
이것을 설명하는 또 다른 방법은 커널 크기 1을 가진 두 개의 컨볼루션이다.
"The dimensionality of input and output is dmodel = 512, and the inner-layer has dimensionalitydf f = 2048"
"입출력의 치수는 dmodel = 512이고, 내부층은 치수 f = 2048이다."

 


3. 4 Embeddings and Softmax

 

Similarly to other sequence transduction models, we use learned embeddings to convert the inputtokens and output tokens to vectors of dimension dmodel"
"임베딩과 소프트맥스 다른 시퀀스 변환 모델과 유사하게, 우리는 입력 토큰과 출력 토큰을 차원 dmodel의 벡터로 변환하기 위해 학습된 임베딩을 사용한다."
We also use the usual learned linear transfor-mation and softmax function to convert the decoder output to predicted next-token probabilities
"또한, 우리는 디코더 출력을 예측된 다음 토큰 확률로 변환하기 위해 일반적인 학습된 선형 변환 및 소프트맥스 함수를 사용한다."
" Inour model, we share the same weight matrix between the two embedding layers and the pre-softmaxlinear transformation, similar to [24]"
Inour 모델에서 우리는 [24]와 유사한 두 개의 임베딩 레이어와 프리 소프트 선형 변환 사이에 동일한 가중치 행렬을 공유합니다.
" In the embedding layers, we multiply those weights by dmodel" ( √ 는 루트다 )
삽입층에서 우리는 그 가중치에 모델 곱하기

 

 

3.5 Positional Encoding

 

Since our model contains no recurrence and no convolution, in order for the model to make use of theorder of the sequence, we must inject some information about the relative or absolute position of the tokens in the sequence"
"5 위치 인코딩우리의 모델이 반복과 무 컨볼루션을 포함하기 때문에, 모델이 시퀀스의 순서를 이용하기 위해서는, 우리는 시퀀스에서 토큰의 상대적 또는 절대적 위치에 대한 정보를 주입해야 한다."



 

 

Table 1: Maximum path lengths, per-layer complexity and minimum number of sequential operationsfor different layer types"
"최대 경로 길이, 계층당 복잡성 및 다른 계층 유형에 대한 최소 순차 연산 수"


" n is the sequence length, d is the representation dimension, k is the kernelsize of convolutions and r the size of the neighborhood in restricted self-attention"
"n은 시퀀스 길이, d는 표현 차원, k는 컨볼루션의 커널 크기, r은 제한된 자기 주의에서 이웃의 크기이다."

 

"To this end, we add ""positional encodings"" to the input embeddings at the bottoms of the encoder and decoder stacks
이를 위해, 우리는 5Table 1의 입력 임베딩에 ""위치 인코딩""을 추가한다. 엔코더 및 디코더 스택의 하단
" The positional encodings have the same dimension dmodelas the embeddings, so that the two can be summed"
위치 인코딩은 임베딩과 동일한 차원 dmodelas를 가지므로 두 가지를 합산할 수 있습니다


" There are many choices of positional encodings,learned and xed [8]"
위치 인코딩에는 많은 선택이 있습니다. 학습 및 고정 [8]


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

 

 

"본 논문에서는 pos가 위치이고 i가 치수인 pos(pos/10002i/dmodel) = sin(pos/100002i/dmodel)P E(pos,2i+1) = cos(pos/10002i/dmodel)의 주파수와 cosine의 함수를 이용하여 pos를 정의하였다."


" That is, each dimension of the positional encoding corresponds to a sinusoid"
", 위치 인코딩의 각 차원은 사인파에 반응한다."


The wavelengths form a geometric progression from 2π to 10000 · 2π
파장은 2에서 10000 · 2까지의 기하학적 진행을 형성한다.


" Wechose this function because we hypothesized it would allow the model to easily learn to attend byrelative positions, since for any xed offset k, P Epos+k can be represented as a linear function ofP Epos"
이 함수를 선택하면 고정 오프셋 k의 경우 P Epos + kP Epos의 선형 함수로 표현 될 수 있기 때문에 모델이 쉽게 상대적 위치에 참석하는 것을 배울 수 있다고 가정했기 때문에 이 함수를 선택하십시오.


"We also experimented with using learned positional embeddings [8] instead, and found that the twoversions produced nearly identical results (see Table 3 row (E))"
또한 학습된 위치 임베딩 [8]을 사용하여 실험을 수행한 결과 두 가지 버전이 거의 동일한 결과를 산출한다는 것을 발견했다(3 (E) 참조).


We chose the sinusoidal versionbecause it may allow the model to extrapolate to sequence lengths longer than the ones encounteredduring training
우리는 정현파 버전을 선택했다. 왜냐하면 그것은 모델이 훈련 중에 마주친 것보다 더 긴 순서 길이로 추론할 수 있게 해줄 수 있기 때문이다.


4 Why Self-Attention

 

In this section we compare various aspects of self-attention layers to the recurrent and convolu-tional layers commonly used for mapping one variable-length sequence of symbol representations(x1, ... , xn) to another sequence of equal length (z1, ... , zn), with xi, zi  Rd, such as a hiddenlayer in a typical sequence transduction encoder or decoder"
"4 왜 자기주의의 이 절에서 우리는 자기주의 계층의 다양한 측면을 하나의 가변 길이 기호 표현 시퀀스 
(x1 , ... , xn)과 동일한 길이의 다른 시퀀스(z1, ..., zn)은 전형적인 시퀀스 트랜스덕션 인코더 또는 디코더의 히든 레이어와 같은 xi, zi Rd를 갖는다."

 


Motivating our use of self-attention weconsider three desiderata
자기주의력 사용 동기 부여 3가지 데시데라타


One is the total computational complexity per layer
하나는 층당 총 계산 복잡성입니다


" Another is the amount of computation that canbe parallelized, as measured by the minimum number of sequential operations required"
또 다른 것은 필요한 최소 순차 연산 수로 측정한 것과 같이 병렬화할 수 있는 계산량입니다


The third is the path length between long-range dependencies in the network
세 번째는 네트워크에서의 장거리 종속성 간의 경로 길이입니다


Learning long-rangedependencies is a key challenge in many sequence transduction tasks
장거리 의존성을 배우는 것은 많은 시퀀스 변환 작업에서 중요한 과제다.


One key factor affecting theability to learn such dependencies is the length of the paths forward and backward signals have totraverse in the network
이러한 의존성을 학습하는 데 영향을 미치는 한 가지 핵심 요소는 순방향 신호와 역방향 신호가 네트워크에서 역행해야 하는 경로 길이입니다


" The shorter these paths between any combination of positions in the inputand output sequences, the easier it is to learn long-range dependencies [11]"

입력 및 출력 시퀀스의 임의의 위치 조합 사이에 이러한 경로가 짧을수록 장거리 종속성을 쉽게 학습할 수 있습니다 [11]


Hence we also comparethe maximum path length between any two input and output positions in networks composed of thedifferent layer types
따라서 우리는 또한 다른 계층 유형으로 구성된 네트워크에서 두 입력 위치와 출력 위치 사이의 최대 경로 길이를 비교한다.


"As noted in Table 1, a self-attention layer connects all positions with a constant number of sequentiallyexecuted operations, whereas a recurrent layer requires O(n) sequential operations"
"1에서 언급한 바와 같이, 자기 주의 계층은 모든 위치를 순차적으로 실행된 일정 수의 연산과 연결하지만, 반복 계층은 O(n) 순차 연산을 요구한다."


" In terms ofcomputational complexity, self-attention layers are faster than recurrent layers when the sequencelength n is smaller than the representation dimensionality d, which is most often the case withsentence representations used by state-of-the-art models in machine translations, such as word-piece[31] and byte-pair [25] representations"
"계산 복잡성 측면에서, 시퀀스 길이 n이 표현 차원 d보다 작을 때 자기 주의 계층은 반복 계층보다 빠르며, 이는 워드 피스 [31] 및 바이트 쌍 [25] 표현과 같은 기계 번역에서 최첨단 모델에서 사용되는 문장 표현의 경우 가장 자주 사용됩니다."


" To improve computational performance for tasks involvingvery long sequences, self-attention could be restricted to considering only a neighborhood of size r in6 the input sequence centered around the respective output position"
"매우 긴 시퀀스를 포함하는 태스크에 대한 계산 성능을 향상시키기 위해, 자기 주의는 각각의 출력 위치를 중심으로 한 입력 시퀀스 r in6 크기의 근방만을 고려하는 것으로 제한될 수 있다."


This would increase the maximumpath length to O(n/r)
이렇게 하면 O(n/r)까지 최대 경로 길이가 증가한다.


We plan to investigate this approach further in future work
우리는 향후 작업에서 이 접근법을 더 조사할 계획이다.


A single convolutional layer with kernel width k < n does not connect all pairs of input and outputpositions
커널 폭 k < n을 갖는 단일 컨볼루션 계층은 모든 쌍의 입력 및 출력을 연결하지 않습니다


" Doing so requires a stack of O(n/k) convolutional layers in the case of contiguous kernels,or O(logk(n)) in the case of dilated convolutions [15], increasing the length of the longest pathsbetween any two positions in the network"
"이렇게 하려면 연속적인 커널의 경우 O(n/k) 컨볼루션 층의 스택이 필요하며, 확장된 컨볼루션의 경우 O(logk(n))가 필요하며, 네트워크의 두 위치 사이에서 가장 긴 경로의 길이를 증가시킨다[15]"


" Convolutional layers are generally more expensive thanrecurrent layers, by a factor of k"
컨볼루션 층은 일반적으로 k의 요인에 의해 반복 층보다 비싸다.


" Separable convolutions [6], however, decrease the complexityconsiderably, to O(k · n · d + n · d2)"
그러나 분리 가능한 컨볼루션[6]은 복잡도를 O(k · n · d + n · d2)로 상당히 감소시킨다.


" Even with k = n, however, the complexity of a separableconvolution is equal to the combination of a self-attention layer and a point-wise feed-forward layer,the approach we take in our model"
그러나 k = n을 사용하더라도 분리 가능한 컨볼루션의 복잡성은 자기 주의 계층과 점별 피드 포워드 계층의 조합과 동일합니다.이 접근법은 모델에서 취합니다


"As side benet, self-attention could yield more interpretable models"
"측면의 이익으로서, 자기 주의는 더 해석 가능한 모델을 산출할 수 있다."


We inspect attention distributionsfrom our models and present and discuss examples in the appendix
우리는 모델에서 주의 분포를 검사하고 부록에 있는 예를 제시하고 토론합니다


" Not only do individual attentionheads clearly learn to perform different tasks, many appear to exhibit behavior related to the syntacticand semantic structure of the sentences"
"개별 주의력은 다른 작업을 수행하는 것을 명확하게 배울 뿐만 아니라, 많은 사람들이 문장의 구문과 의미 구조와 관련된 행동을 보이는 것처럼 보인다."


5 Training

 

This section describes the training regime for our models
5 훈련이 절에서는 우리 모델의 훈련 체제를 설명한다.

5. 1 Training Data and Batching

 

We trained on the standard WMT 2014 English-German dataset consisting of about 4.5 million sentence pairs
우리는 약 4500만 쌍 문장으로 구성된 표준 WMT 2014 영어-독일 데이터 세트를 훈련했다.

" Sentences were encoded using byte-pair encoding [3], which has a shared source-target vocabulary of about 37000 tokens"
문장은 약 37000개의 토큰의 공유 소스 - 타겟 어휘를 갖는 바이트 쌍 인코딩 [3]을 사용하여 인코딩되었습니다.


" For English-French, we used the signicantly larger WMT2014 English-French dataset consisting of 36M sentences and split tokens into a 32000 word-piecevocabulary [31]"

영어-프랑스어의 경우 36M 문장과 분할 토큰으로 구성된 상당히 큰 WMT2014 영어-프랑스어 데이터 세트를 32000 단어-피스 어휘로 사용했다 [31]


Sentence pairs were batched together by approximate sequence length
문장 쌍은 대략적인 시퀀스 길이에 의해 함께 배치되었다.


Each trainingbatch contained a set of sentence pairs containing approximately 25000 source tokens and 25000target tokens
"각 훈련 배치에는 약 25,000개의 소스 토큰과 25,000개의 타겟 토큰이 포함된 문장 쌍이 포함되어 있었다."


5. 2 Hardware and Schedule

 

We trained our models on one machine with 8 NVIDIA P100 GPUs
2 하드웨어 및 스케줄 우리는 8개의 NVIDIA P100 GPU로 하나의 기계에서 모델을 훈련했다.
" For our base models usingthe hyperparameters described throughout the paper, each training step took about 0.4 seconds"
논문 전체에 설명된 하이퍼 파라미터를 사용하는 기본 모델의 경우 각 훈련 단계는 약 0.4 초가 걸렸습니다


" Wetrained the base models for a total of 100,000 steps or 12 hours"
기본 모델을 총 10만 단계 또는 12시간 동안 젖혔습니다


" For our big models,(described on thebottom line of table 3), step time was 1.0 seconds"
"우리의 큰 모델의 경우, (3의 하단 라인에 설명) 단계 시간은 1이었다."


" The big models were trained for 300,000 steps(3.5 days)
이 대형 모델들은 30만 단계로 훈련되었다(3 .5)

 

5.3 Optimizer

 

We used the Adam optimizer [17] with β1 = 0.9, β2 = 0.98 ,and e = 10^-9 
우리는 β1 = 0.9, β2 = 0.98 , e = 10^-9 을 가진 Adam 최적화 [17]를 사용했다.

We varied the learningrate over the course of training, according to the formula:

우리는 훈련 과정에서 학습률을 변화시켰다


This corresponds to increasing the learning rate linearly for the rst warmup_steps training steps,and decreasing it thereafter proportionally to the inverse square root of the step number"
"5)(3) 이는 제 1 워밍업_스텝 훈련 단계에 대해 선형적으로 학습 속도를 증가시키고, 그 후 스텝 수의 역제곱근에 비례하여 학습 속도를 감소시키는 것에 해당한다."


We usedwarmup_steps = 4000
우리는 warmup_steps = 4000을 사용했다.

 


5.4 Regularization

 

We employ three types of regularization during training:

정규화 훈련 중 정규화는 세 가지 유형을 사용합니다. 

 

Residual Dropout We apply dropout [27] to the output of each sub-layer, before it is added to thesub-layer input and normalized"


잔류 중퇴 우리는 하위 계층 입력에 추가되고 정규화되기 전에 각 하위 계층의 출력에 중퇴 [27]를 적용합니다


" In addition, we apply dropout to the sums of the embeddings and thepositional encodings in both the encoder and decoder stacks"
"또한, 인코더 및 디코더 스택의 임베딩 및 위치 인코딩의 합에 드롭아웃을 적용합니다"


" For the base model, we use a rate of Pdrop = 0.1"
"기본 모델의 경우, 우리는 Pdrop = 0.1 의 비율을 사용한다."

7Table 2: The Transformer achieves better BLEU scores than previous state-of-the-art models on theEnglish-to-German and English-to-French newstest2014 tests at a fraction of the training cost
72: 트랜스포머는 훈련 비용의 극히 일부에서 영어 대 독일 및 영어 대 프랑스 뉴스테스트2014 테스트에서 이전의 최첨단 모델보다 더 나은 BLEU 점수를 획득한다.

 

 

 

Label Smoothing During training, we employed label smoothing of value e ls = 0.1 [30]
라벨 평활화 훈련 중, 우리는 값 평활화 = 0의 라벨 평활화를 사용하였다."


" Thishurts perplexity, as the model learns to be more unsure, but improves accuracy and BLEU score"
이 모델은 더 확신하지 못하는 것을 배우기 때문에 당혹감을 주지만 정확성과 BLEU 점수를 향상시킨다.

 


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"
1 WMT 2014 영어-독일 번역 작업에서 큰 변압기 모델(Transformer (big)in Table 2)은 이전에 보고된 가장 좋은 모델( 앙상블 포함)2 이상 능가한다.


"0BLEU, establishing a new state-of-the-art BLEU score of 28.4 
"0BLEU, 28.4 점의 새로운 최첨단 BLEU 점수 수립"

The conguration of this model islisted in the bottom line of Table 3
이 모델의 구성은 표 3의 맨 아래 선에 나열되어 있습니다
Training took 3
훈련은 3이 걸렸다
5 days on 8 P100 GPUs
8 P100 GPU 5
" Even our base modelsurpasses all previously published models and ensembles, at a fraction of the training cost of any ofthe competitive models"
심지어 우리의 기본 모델도 이전에 발표된 모든 모델과 앙상블을 경쟁 모델의 훈련 비용의 일부로 능가한다.


"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 theprevious state-of-the-art model""
WMT 2014 영어-프랑스어 번역 과제에서 우리의 큰 모델은 41BLEU 점수를 달성한다 이전에 발표된 모든 단일 모델보다 성능이 우수하며, 이전 최첨단 모델의 훈련 비용의 1/4 미만이다.


The Transformer (big) model trained for English-to-French useddropout rate Pdrop = 0.1, instead of 0.3
영어 대 프랑스어 중고 드롭아웃 속도 Pdrop = 0.1 대신 3.3 으로 훈련된 트랜스포머() 모델

"For the base models, we used a single model obtained by averaging the last 5 checkpoints, which were written at 10-minute intervals"
"기본 모델의 경우, 마지막 5개의 체크포인트를 평균하여 얻은 단일 모델을 사용했는데, 이 모델은 10분 간격으로 작성되었다."


" For the big models, we averaged the last 20 checkpoints"
"대형 모델의 경우, 우리는 지난 20개의 검문소를 평균적으로 계산했습니다"


Weused beam search with a beam size of 4 and length penalty α = 0.6
4와 길이 페널티 α = 0.6 의 빔 크기를 가진 Weused 빔 검색

These hyperparameterswere chosen after experimentation on the development set
이러한 하이퍼 파라미터는 개발 세트에 대한 실험 후에 선택되었다.


" We set the maximum output length duringinference to input length + 50, but terminate early when possible [31]"
우리는 입력 길이 + 50으로 추론하는 동안 최대 출력 길이를 설정하지만 가능하면 일찍 종료합니다 [31]


Table 2 summarizes our results and compares our translation quality and training costs to other modelarchitectures from the literature
2는 우리의 결과를 요약하고 우리의 번역 품질과 훈련 비용을 문헌의 다른 모델 구조와 비교한다.


" We estimate the number of oating point operations used to train amodel by multiplying the training time, the number of GPUs used, and an estimate of the sustainedsingle-precision oating-point capacity of each GPU (5)"
"우리는 훈련 시간, 사용 GPU , GPU (5)의 지속 정밀 부동 소수점 용량의 추정치를 곱하여 모델 훈련에 사용되는 부동 소수점 연산 수를 추정한다."

 

6.2  Model Variations

 

To evaluate the importance of different components of the Transformer, we varied our base modelin different ways, measuring the change in performance on English-to-German translation on thedevelopment set, newstest2013"
"2 모델 변이를 평가하기 위해, 우리는 개발 세트, 뉴스테스트2013에서 영어-독일 번역의 성능 변화를 측정하면서, 트랜스포머의 다른 구성 요소의 중요성을 평가하기 위해, 우리의 기본 모델을 다른 방식으로 변화시켰다."


" We used beam search as described in the previous section, but nocheckpoint averaging"
"이전 섹션에서 설명한 대로 빔 검색을 사용했지만, 노체크포인트 평균"
We present these results in Table 3
이 결과를 표 3에 제시합니다


"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"
"3(A)에서, 우리는 제3절에서 설명한 대로 계산의 양을 일정하게 유지하면서 주의 헤드의 수와 주의 키 및 값 치수를 변경한다."


While single-headattention is 0.9 BLEU worse than the best setting, quality also drops off with too many heads"
단일 머리 주의가 0.9 BLEU가 가장 좋은 설정보다 더 나쁘고, 품질도 너무 많은 머리로 떨어진다."


(5)We used values of 2.8, 3.7, 6.0 and 9.5 TFLOPS for K80, K40, M40 and P100, respectively

K80, K40, M40, P100은 각각 2.8, 3.7, 6.0, 9.5 TFLOPS의 값을 사용하였다.


8 Table 3: Variations on the Transformer architecture
83: 변압기 구조의 변화
Unlisted values are identical to those of the basemodel
목록에 없는 값은 기준 모델의 값과 동일합니다
" All metrics are on the English-to-German translation development set, newstest2013"
"모든 측정 기준은 영어 대 독일 번역 개발 세트, 뉴스테스트2013에 있다."
" Listedperplexities are per-wordpiece, according to our byte-pair encoding, and should not be compared toper-word perplexities"
목록 복잡성은 바이트 쌍 인코딩에 따라 단어 당이며 단어 당 복잡성을 비교해서는 안됩니다.

 

 

In Table 3 rows (B), we observe that reducing the attention key size dk hurts model quality"
3(B)에서 주의 키 크기 dk를 줄이면 모델 품질이 저하되는 것을 관찰한다.


Thissuggests that determining compatibility is not easy and that a more sophisticated compatibilityfunction than dot product may be benecial
이는 호환성을 결정하는 것이 쉽지 않으며 점 제품보다 더 정교한 호환성 기능이 유익할 수 있음을 시사한다.


" We further observe in rows (C) and (D) that, as expected,bigger models are better, and dropout is very helpful in avoiding over-tting"
"우리는 예상대로 더 큰 모델이 더 좋고, 중퇴가 과적합을 피하는 데 매우 도움이 된다는 것을 행(C)(D)에서 더 관찰한다."


" In row (E) we replace oursinusoidal positional encoding with learned positional embeddings [8], and observe nearly identicalresults to the base model"
"로우 (E)에서 우리는 정현파 위치 인코딩을 학습된 위치 임베딩으로 대체하고, 기본 모델과 거의 동일한 결과를 관찰한다."


7 Conclusion

 

In this work, we presented the Transformer, the rst sequence transduction model based entirely onattention, replacing the recurrent layers most commonly used in encoder-decoder architectures withmulti-headed self-attention"
"본 논문에서는 인코더-디코더 아키텍처에서 가장 많이 사용되는 반복 레이어를 다중 헤드 자기 주의로 대체하여, 전적으로 주의를 기반으로 한 첫 번째 시퀀스 변환 모델인 트랜스포머를 제시하였다."


"For translation tasks, the Transformer can be trained signicantly faster than architectures basedon recurrent or convolutional layers"
변환 작업의 경우 변환기는 반복 또는 컨볼루션 계층을 기반으로 하는 아키텍처보다 훨씬 빠르게 훈련될 수 있다.


" On both WMT 2014 English-to-German and WMT 2014English-to-French translation tasks, we achieve a new state of the art"
2014WMT 영어-독일어와 WMT 2014년 영어-프랑스어 번역 작업 모두에서 우리는 새로운 예술 상태를 달성한다.


In the former task our bestmodel outperforms even all previously reported ensembles
이전 작업에서 우리의 베스트모델은 이전에 보고된 모든 앙상블보다 더 잘 작동한다.


We are excited about the future of attention-based models and plan to apply them to other tasks
우리는 관심 기반 모델의 미래에 대해 흥분하고 있으며 다른 작업에 적용할 계획이다


" Weplan to extend the Transformer to problems involving input and output modalities other than text andto investigate local, restricted attention mechanisms to efciently handle large inputs and outputssuch as images, audio and video"
"변환기를 텍스트 이외의 입력 및 출력 양식과 관련된 문제로 확장하고 이미지, 오디오 및 비디오와 같은 큰 입력 및 출력을 효율적으로 처리하기 위한 로컬 제한 주의 메커니즘을 조사할 계획"


Making generation less sequential is another research goals of ours
세대를 덜 순차적으로 만드는 것은 우리의 또 다른 연구 목표다.


The code we used to train and evaluate our models is available at https://github.com/tensorflow/tensor2tensor
우리가 모델들을 훈련시키고 평가하는 데 사용한 코드는 깃헙에서 구할 수 있다.


Acknowledgements

We are grateful to Nal Kalchbrenner and Stephan Gouws for their fruitfulcomments, corrections and inspiration"
"우리는 Nal KalchbrennerStephan Gouws가 그들의 결실, 수정, 영감을 준 것에 대해 감사한다."

 

 

레퍼런스

[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint 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.

[3] Denny Britz, Anna Goldie, Minh-Thang Luong, and Quoc V. Le. Massive exploration of neural machine translation architectures. 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.

[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.

[6] Francois Chollet. Xception: Deep learning with depthwise separable convolutions. arXiv preprint 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.

[8] Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. Convolutional sequence to sequence learning. arXiv preprint arXiv:1705.03122v2, 2017.

[9] Alex Graves. Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850, 2013.

[10] 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.

[11] Sepp Hochreiter, Yoshua Bengio, Paolo Frasconi, and Jürgen Schmidhuber. Gradient flow in recurrent nets: the difficulty of learning long-term dependencies, 2001.

[12] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997.

[13] Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. Exploring the limits of language modeling. arXiv preprint arXiv:1602.02410, 2016.

[14] Łukasz Kaiser and Ilya Sutskever. Neural GPUs learn algorithms. In International Conference on Learning Representations (ICLR), 2016.

[15] 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.

[16] Yoon Kim, Carl Denton, Luong Hoang, and Alexander M. Rush. Structured attention networks. In International Conference on Learning Representations, 2017.

[17] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015.

[18] Oleksii Kuchaiev and Boris Ginsburg. Factorization tricks for LSTM networks. arXiv preprint arXiv:1703.10722, 2017.

[19] 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.

[20] Samy Bengio Łukasz Kaiser. Can active memory replace attention? In Advances in Neural Information Processing Systems, (NIPS), 2016. 10

[21] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attentionbased neural machine translation. arXiv preprint arXiv:1508.04025, 2015.

[22] Ankur Parikh, Oscar Täckström, Dipanjan Das, and Jakob Uszkoreit. A decomposable attention model. In Empirical Methods in Natural Language Processing, 2016.

[23] Romain Paulus, Caiming Xiong, and Richard Socher. A deep reinforced model for abstractive summarization. arXiv preprint arXiv:1705.04304, 2017.

[24] Ofir Press and Lior Wolf. Using the output embedding to improve language models. arXiv preprint arXiv:1608.05859, 2016.

[25] Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015.

[26] 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.

[27] 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.

[28] 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.

[29] 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.

[30] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. CoRR, abs/1512.00567, 2015.

[31] 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.

[32] 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.

반응형

댓글