SlideShare a Scribd company logo
1
DEEP LEARNING JP
[DL Papers]
http://deeplearning.jp/
“Training RNNs as Fast as CNNs”
Hiroki Kurotaki, Matsuo Lab
目目次次
• 概要
• 1 Introduction
• 2 Method
• 3 Related Works
• 4 Experiments
• 5 Conclusion
2
目目次次
• 概要
• 1 Introduction
• 2 Method
• 3 Related Works
• 4 Experiments
• 5 Conclusion
3
書書誌誌情情報報
• Training RNNs as Fast as CNNs
• Tao Lei, Yu Zhang
• 9/12/2017(v1: 9/8/2017)
• https://arxiv.org/abs/1709.02755v2
• https://github.com/taolei87/sru
• Arxiv Sanityで Last monthのtop hype #2 (329 tweets)
• 1st authorはICML, EMNLPなどに通している
– Deriving neural architectures from sequence and graph kernels
4
提提案案手手法法
• RNNのゲートに前の時間の情報を入れない
– 大幅な並列化が可能
• cuDNNで最適化されたLSTMに比べ、5-10x 速い
• PyTorch and CNTKでオープンソース公開
5
主主なな結結果果
• 平均実行時間の比較
• CuDNNのLSTM実装より10倍速い
6
目目次次
• 概要
• 1 Introduction
• 2 Method
• 3 Related Works
• 4 Experiments
• 5 Conclusion
7
11 IInnttrroodduuccttiioonn
• 深層学習の研究開発において、実行時間は大きな障害
• LSTMは並列化の恩恵を最大限に受け取れていない
– h_tがh_{t-1}に依存しているため、並列化が不可能
• 依存項をカットした、Simple Recurrent Unitを提案
• CUDAレベルで最適化した実装を公開した
– conv2dと同等の速度を達成した
8
目目次次
• 概要
• 1 Introduction
• 2 Method
• 3 Related Works
• 4 Experiments
• 5 Conclusion
9
22..11 SSRRUU iimmpplleemmeennttaattiioonn
• ベース:LSTM+頻出のテクニック二つ
– Highway connection
• 下のh’_tの式。r_tがreset gateと呼ばれるもの
– Variational dropout :
• 入力x_tに時間で変わらないマスク
• 細かいこと
– Forget gateは、i = 1-fとする
– h_tに普通のdropout
– g(・)は活性化関数
10
22..22 SSppeeeeddiinngg--uupp tthhee rreeccuurrrreennccee
• 従来のボトルネック
– 隠れ状態の各次元が、他を参照してしまい、並列化が不可能
– h_{t-1}の全体が計算されるまで待たないと、h_tを計算不可
• 提案: ゲートにおける時間t-1の参照をカット
– ボトルネックは(3)-(5)の行列計算のみ
11
22..33 CCUUDDAA lleevveell ooppttiimmiizzaattiioonn
• (3)-(5)式の行列演算は一つにまとめる
12
22..33 CCUUDDAA lleevveell ooppttiimmiizzaattiioonn
• 計算が並列化できるようになる
13
目目次次
• 概要
• 1 Introduction
• 2 Method
• 3 Related Works
• 4 Experiments
• 5 Conclusion
14
33 RReellaatteedd WWoorrkk
• 系列処理の効率化
– Recurrent convolution (RCNN) (Lei et al., 2015, 2016)
– kernel network (KNN) (Lei et al., 2017)
– Quasi-RNN (Bradbury et al., 2017)
• カットによる表現力の減少有無
– 単純化RNNのcapacityの調査(Balduzzi and Ghifary (2016))
– SRUやword-level CNNは、系列類似度関数→隠れ空間の埋め込み
(Lei et al. (2017))
15
目目次次
• 概要
• 1 Introduction
• 2 Method
• 3 Related Works
• 4 Experiments
• 5 Conclusion
16
44 EExxppeerriimmeennttss
• 提案手法のSRUを、先行研究やCuDNNのLSTM実装と比較
• SRUの、レイヤーを積み増すバージョンで、
良い精度と速度を出した
• 実装は4.5以外PyTorch、4.5はCNTK
17
44..11 CCllaassssiiffiiccaattiioonn
• データセット
– movie reviews (MR) (Pang and Lee, 2005)
– subjectivity data (SUBJ) (Pang and Lee, 2004)
– customer reviews (CR) (Hu and Liu, 2004)
– TREC questions (Li and Roth, 2002)
– opinion polarity from MPQA data (Wiebe et al., 2005)
– Stanford sentiment treebank (SST) (Socher et al., 2013)
• モデル、準備
– 2レイヤー、隠れ128次元
• SSTデータセットでは4レイヤー
– CNNでも比較
• (Convolutional neural networks for sentence classification)
18
44..11 CCllaassssiiffiiccaattiioonn
• 良い結果と速度が出た
19
44..11 CCllaassssiiffiiccaattiioonn
20
44..22 QQuueessttiioonn aannsswweerriinngg
• データセット
– Stanford Question Answering Dataset (SQuAD) (Rajpurkar et al.,
2016)
• wikipedia からの100,000 QAペア
• モデル、準備
– Document Reader model (Chen et al., 2017)
• LSTM版とSRU版を作って比較
– 50エポック、32バッチ、隠れ128次元、
– ドロップアウト 入力0.5、SRU0.2、LSTM0.3
21
44..22 QQuueessttiioonn aannsswweerriinngg
• LSTMは69.6%マッチ、78.9% F1スコア
• SRUは70.3%マッチ、79.5% F1スコア
• 6~10倍の高速化
22
44..33 LLaanngguuaaggee mmooddeelliinngg
• データセット
– Penn Treebank corpus (PTB)
• 1Mトークン、10k辞書
• truncated BPTTで学習
• モデル、前準備
– truncated BPTTが35エポック、バッチサイズ32、dropout0.75
– 300エポックの訓練
23
44..33 LLaanngguuaaggee mmooddeelliinngg
• Perplexitiesで先行研究やcuDNN LSTMを上回る
24
44..44 MMaacchhiinnee ttrraannssllaattiioonn
• データセット
– WMT’14 English→German translation
– 4Mの翻訳ペア
• モデル、前処理
– OpenNMTという翻訳システムをSRUに拡張した
– seq2seq w/ attention
• h_{t-1}は並列化を妨げるため、次の時間の入力には追加しない
– 15エポック、バッチ64、word embeddings size 500
– dropout rateを、よく使われるものより小さい0.1に落とした
25
44..44 MMaacchhiinnee ttrraannssllaattiioonn
• BLEUスコアで、元論文を上回る
26
44..55 SSppeeeecchh rreeccooggnniittiioonn
• データセット
– Switchboard-1 corpus (Godfrey et al., 1992)
• 4,870会話(300時間) 話者520人
• モデルなど
– MFCC、Kaldiを使用
– Computational Network Toolkit (CNTK)で実装
27
44..55 SSppeeeecchh rreeccooggnniittiioonn
• SOTAの結果
28
目目次次
• 概要
• 1 Introduction
• 2 Method
• 3 Related Works
• 4 Experiments
• 5 Conclusion
29
55 CCoonncclluussiioonn
• Simple Recurrent Unit (SRU)を提案
– ゲートのh_{t-1}参照項をカット
• 5つのタスクで性能を確認した
• 従来のCuDNNのLSTM実装などに比べ、最大10倍の高速化
– 精度も向上した
30
31

More Related Content

PDF
[DL輪読会] Spectral Norm Regularization for Improving the Generalizability of De...
PDF
[DLHacks 実装] DeepPose: Human Pose Estimation via Deep Neural Networks
PDF
[DL輪読会] Towards an Automatic Turing Test: Learning to Evaluate Dialogue Respo...
PDF
論文 Solo Advent Calendar
PPTX
[DL輪読会]YOLOv4: Optimal Speed and Accuracy of Object Detection
PDF
[DLHacks 実装]Network Dissection: Quantifying Interpretability of Deep Visual R...
PDF
Deep Learning技術の最近の動向とPreferred Networksの取り組み
PDF
[DL輪読会]The Neural Process Family−Neural Processes関連の実装を読んで動かしてみる−
[DL輪読会] Spectral Norm Regularization for Improving the Generalizability of De...
[DLHacks 実装] DeepPose: Human Pose Estimation via Deep Neural Networks
[DL輪読会] Towards an Automatic Turing Test: Learning to Evaluate Dialogue Respo...
論文 Solo Advent Calendar
[DL輪読会]YOLOv4: Optimal Speed and Accuracy of Object Detection
[DLHacks 実装]Network Dissection: Quantifying Interpretability of Deep Visual R...
Deep Learning技術の最近の動向とPreferred Networksの取り組み
[DL輪読会]The Neural Process Family−Neural Processes関連の実装を読んで動かしてみる−

What's hot (20)

PPTX
Res netと派生研究の紹介
PPTX
[DL輪読会]Meta-Learning Probabilistic Inference for Prediction
PDF
第1回NIPS読み会・関西発表資料
PPTX
[DL輪読会]LightTrack: A Generic Framework for Online Top-Down Human Pose Tracking
PPTX
Pred net使ってみた
PDF
ArtTrack: Articulated Multi-Person Tracking in the Wild : CV勉強会関東
PPTX
MS COCO Dataset Introduction
PPTX
[DL輪読会]Learning to Adapt: Meta-Learning for Model-Based Control
PDF
SSD: Single Shot MultiBox Detector (ECCV2016)
PDF
[DL輪読会]Auto-DeepLab: Hierarchical Neural Architecture Search for Semantic Ima...
PDF
0から理解するニューラルネットアーキテクチャサーチ(NAS)
PPTX
[DL輪読会]DropBlock: A regularization method for convolutional networks
PPTX
[DL輪読会]Deep Face Recognition: A Survey
PPTX
[DL輪読会]Objects as Points
PDF
Overcoming Catastrophic Forgetting in Neural Networks読んだ
PPTX
[DL輪読会]Xception: Deep Learning with Depthwise Separable Convolutions
PPTX
[DL輪読会]SOLAR: Deep Structured Representations for Model-Based Reinforcement L...
PDF
論文紹介 "DARTS: Differentiable Architecture Search"
PDF
CVPR2015読み会 "Joint Tracking and Segmentation of Multiple Targets"
PDF
Deep Residual Learning (ILSVRC2015 winner)
Res netと派生研究の紹介
[DL輪読会]Meta-Learning Probabilistic Inference for Prediction
第1回NIPS読み会・関西発表資料
[DL輪読会]LightTrack: A Generic Framework for Online Top-Down Human Pose Tracking
Pred net使ってみた
ArtTrack: Articulated Multi-Person Tracking in the Wild : CV勉強会関東
MS COCO Dataset Introduction
[DL輪読会]Learning to Adapt: Meta-Learning for Model-Based Control
SSD: Single Shot MultiBox Detector (ECCV2016)
[DL輪読会]Auto-DeepLab: Hierarchical Neural Architecture Search for Semantic Ima...
0から理解するニューラルネットアーキテクチャサーチ(NAS)
[DL輪読会]DropBlock: A regularization method for convolutional networks
[DL輪読会]Deep Face Recognition: A Survey
[DL輪読会]Objects as Points
Overcoming Catastrophic Forgetting in Neural Networks読んだ
[DL輪読会]Xception: Deep Learning with Depthwise Separable Convolutions
[DL輪読会]SOLAR: Deep Structured Representations for Model-Based Reinforcement L...
論文紹介 "DARTS: Differentiable Architecture Search"
CVPR2015読み会 "Joint Tracking and Segmentation of Multiple Targets"
Deep Residual Learning (ILSVRC2015 winner)
Ad

Viewers also liked (8)

PPTX
[DL Hacks 実装]The Conditional Analogy GAN: Swapping Fashion Articles on People...
PDF
[DLHacks 実装]Neural Machine Translation in Linear Time
PDF
[DL輪読会]Deep Recurrent Generative Decoder For Abstractive Text Summarization(E...
PDF
[DLHacks 実装]Perceptual Adversarial Networks for Image-to-Image Transformation
PDF
[DL輪読会]Energy-based generative adversarial networks
PDF
[DL輪読会]Opening the Black Box of Deep Neural Networks via Information
PDF
[DL輪読会]Learning to Act by Predicting the Future
PDF
[DL輪読会]Deep Direct Reinforcement Learning for Financial Signal Representation...
[DL Hacks 実装]The Conditional Analogy GAN: Swapping Fashion Articles on People...
[DLHacks 実装]Neural Machine Translation in Linear Time
[DL輪読会]Deep Recurrent Generative Decoder For Abstractive Text Summarization(E...
[DLHacks 実装]Perceptual Adversarial Networks for Image-to-Image Transformation
[DL輪読会]Energy-based generative adversarial networks
[DL輪読会]Opening the Black Box of Deep Neural Networks via Information
[DL輪読会]Learning to Act by Predicting the Future
[DL輪読会]Deep Direct Reinforcement Learning for Financial Signal Representation...
Ad

Similar to [DL輪読会]Training RNNs as Fast as CNNs (20)

PPTX
[DL輪読会]QUASI-RECURRENT NEURAL NETWORKS
PDF
Tree-to-Sequence Attentional Neural Machine Translation (ACL 2016)
PDF
[DL輪読会]QUASI-RECURRENT NEURAL NETWORKS
PPTX
NeurIPS2019参加報告
PDF
Twitterにおける即時話題推定技術「どたばたかいぎ」の開発
PDF
A PID Controller Approach for Stochastic Optimization of Deep Networks
PDF
Development and Experiment of Deep Learning with Caffe and maf
PDF
経済学のための実践的データ分析 4.SQL ことはじめ
PDF
ディープラーニング最近の発展とビジネス応用への課題
PDF
Deep learning reading club @ nimiri for SWEST
PDF
Qaシステム解説
PDF
研究室輪読 Feature Learning for Activity Recognition in Ubiquitous Computing
PPTX
[DL輪読会]ODT: Online Decision Transformer
PPTX
Generating Better Search Engine Text Advertisements with Deep Reinforcement L...
PPTX
関西CVPRML勉強会資料20150627
PDF
GTC 2018 で発表されたディープラーニング最新情報
PDF
深層学習(岡本孝之 著) - Deep Learning chap.1 and 2
PPTX
Multi Scale Recognition with DAG-CNNs
PDF
Survey of Scientific Publication Analysis by NLP and CV
[DL輪読会]QUASI-RECURRENT NEURAL NETWORKS
Tree-to-Sequence Attentional Neural Machine Translation (ACL 2016)
[DL輪読会]QUASI-RECURRENT NEURAL NETWORKS
NeurIPS2019参加報告
Twitterにおける即時話題推定技術「どたばたかいぎ」の開発
A PID Controller Approach for Stochastic Optimization of Deep Networks
Development and Experiment of Deep Learning with Caffe and maf
経済学のための実践的データ分析 4.SQL ことはじめ
ディープラーニング最近の発展とビジネス応用への課題
Deep learning reading club @ nimiri for SWEST
Qaシステム解説
研究室輪読 Feature Learning for Activity Recognition in Ubiquitous Computing
[DL輪読会]ODT: Online Decision Transformer
Generating Better Search Engine Text Advertisements with Deep Reinforcement L...
関西CVPRML勉強会資料20150627
GTC 2018 で発表されたディープラーニング最新情報
深層学習(岡本孝之 著) - Deep Learning chap.1 and 2
Multi Scale Recognition with DAG-CNNs
Survey of Scientific Publication Analysis by NLP and CV

More from Deep Learning JP (20)

PPTX
【DL輪読会】AdaptDiffuser: Diffusion Models as Adaptive Self-evolving Planners
PPTX
【DL輪読会】事前学習用データセットについて
PPTX
【DL輪読会】 "Learning to render novel views from wide-baseline stereo pairs." CVP...
PPTX
【DL輪読会】Zero-Shot Dual-Lens Super-Resolution
PPTX
【DL輪読会】BloombergGPT: A Large Language Model for Finance arxiv
PPTX
【DL輪読会】マルチモーダル LLM
PDF
【 DL輪読会】ToolLLM: Facilitating Large Language Models to Master 16000+ Real-wo...
PPTX
【DL輪読会】AnyLoc: Towards Universal Visual Place Recognition
PDF
【DL輪読会】Can Neural Network Memorization Be Localized?
PPTX
【DL輪読会】Hopfield network 関連研究について
PPTX
【DL輪読会】SimPer: Simple self-supervised learning of periodic targets( ICLR 2023 )
PDF
【DL輪読会】RLCD: Reinforcement Learning from Contrast Distillation for Language M...
PDF
【DL輪読会】"Secrets of RLHF in Large Language Models Part I: PPO"
PPTX
【DL輪読会】"Language Instructed Reinforcement Learning for Human-AI Coordination "
PPTX
【DL輪読会】Llama 2: Open Foundation and Fine-Tuned Chat Models
PDF
【DL輪読会】"Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware"
PPTX
【DL輪読会】Parameter is Not All You Need:Starting from Non-Parametric Networks fo...
PDF
【DL輪読会】Drag Your GAN: Interactive Point-based Manipulation on the Generative ...
PDF
【DL輪読会】Self-Supervised Learning from Images with a Joint-Embedding Predictive...
PPTX
【DL輪読会】Towards Understanding Ensemble, Knowledge Distillation and Self-Distil...
【DL輪読会】AdaptDiffuser: Diffusion Models as Adaptive Self-evolving Planners
【DL輪読会】事前学習用データセットについて
【DL輪読会】 "Learning to render novel views from wide-baseline stereo pairs." CVP...
【DL輪読会】Zero-Shot Dual-Lens Super-Resolution
【DL輪読会】BloombergGPT: A Large Language Model for Finance arxiv
【DL輪読会】マルチモーダル LLM
【 DL輪読会】ToolLLM: Facilitating Large Language Models to Master 16000+ Real-wo...
【DL輪読会】AnyLoc: Towards Universal Visual Place Recognition
【DL輪読会】Can Neural Network Memorization Be Localized?
【DL輪読会】Hopfield network 関連研究について
【DL輪読会】SimPer: Simple self-supervised learning of periodic targets( ICLR 2023 )
【DL輪読会】RLCD: Reinforcement Learning from Contrast Distillation for Language M...
【DL輪読会】"Secrets of RLHF in Large Language Models Part I: PPO"
【DL輪読会】"Language Instructed Reinforcement Learning for Human-AI Coordination "
【DL輪読会】Llama 2: Open Foundation and Fine-Tuned Chat Models
【DL輪読会】"Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware"
【DL輪読会】Parameter is Not All You Need:Starting from Non-Parametric Networks fo...
【DL輪読会】Drag Your GAN: Interactive Point-based Manipulation on the Generative ...
【DL輪読会】Self-Supervised Learning from Images with a Joint-Embedding Predictive...
【DL輪読会】Towards Understanding Ensemble, Knowledge Distillation and Self-Distil...

[DL輪読会]Training RNNs as Fast as CNNs