際際滷

際際滷Share a Scribd company logo
Learning Transferable Architectures
for Scalable Image Recognition
Barret Zoph, Vijay Vasudevanm Jonathon Shlens, Quoc V. Le
Google Brain
覈谿
 Introduction
 Contribution
 Method
 Result
Introduction
 Neural Architecture Search with Reinforcement Learning(NAS)  朱
 NAS : CNN(CIFAR-10), RNN(Penn Treebank) 蟲譟 り
 CIFAR-10 企朱  一危一 旧 800 GPU, 28days    一危一??
 Learning Transferable Architectures for Scalable Image Recognition (2018, CVPR)
 CNN 豐 襷豢 伎 螳ロ螻   architecture襯 (NASNet)
 CIFAR-10朱 谿場 Convolution Cell 伎 ImageNet   SOTA 焔 !
 NAS 觜 旧  螳 豢 (500GPU, 4days , x7 speed up)
* NAS : Nvidia K40 GPU / NASNet : Nvidia P100s GPU
Contribution(1)
  Search Space
 ろ語 豌企ゼ 覯 り(NAS)  Normal Cell, Reduction Cell 2螳 Convolution Cell 牛 り
 Block  Cell  Architecture
Block
Contribution(2)
 Transferability & Good Performance
 CIFAR-10朱 谿場 Convolution Cell 伎   Dataset  螳
 CIFAR-10, ImageNet 蠍一 SOTA 焔 
 Object Detection NASNet 蟲譟磯ゼ    焔
Method(1)
 Overall architectures of the convolutional nets are manually
predetermined
 Normal Cell - convolutional cells that return a feature map of the same
dimension
 Reduction Cell - convolutional cells that return a feature map where the
feature map height and width is reduced by a factor of two (Initial
operation with stride of two conv)
 Using common heuristic to double the number of filters in the
output whenever the spatial activation size is reduced
 N, 豌 convolution cell filter 螳 User螳 讌伎殊伎  覲
Method(2)
 Overall architecture 蟲
 Controller(LSTM)襯 伎 Normal Cell, Reduction Cell 
 1 Block = hidden state input 2螳 + operation 2螳 + combine(add or concat) 1螳 豐 5螳 覓語
 1螳 Cell  B螳 Block朱 蟲 (朱語 B=5  焔 一)
 Normal Cell 企 predictions 5B螳 + Reduction Cell 企 predictions 5B螳
Block
Method(2)
 Overall architecture 蟲 (Cont.)
 Controller(LSTM)襯 伎 Normal Cell, Reduction Cell 
 1 Block = hidden state input 2螳 + operation 2螳 + combine(add or concat) 1螳 豐 5螳 覓語
 Operation 譬襯  れ 覦覯襦 譴 13螳
 stride = 1 (Normal Cell) / stride = 1 or 2 (Reduction Cell)
Method(2) - Supplement
 Overall architecture 蟲(Cont.)
 stride = 1 or 2 (Reduction Cell)
(豢豌 : https://github.com/titu1994/Keras-NASNet/blob/master/nasnet.py)
Method(2)  NASNet-A
揃 揃 揃
<Normal Cell>
揃 揃 揃
<Reduction Cell>
B螳
B螳
Training with RL
 Controller(LSTM)襯 伎 豸″ Cell 牛 Architecture 
 螳 螻襴讀朱 REINFORCE  Proximal Policy Optimization(PPO) 
 PPO : 2017, OpenAI
 豌伎  覦覯 NAS 蟇一 
 State : controller hidden state
 Action : controller螳 燕 predictions
 Reward : Overall architecture accuracy
Result(CIFAR-10)
 蠍一ヾ SOTA襦 れ network 觜
一 焔!
 NAS CNN architecture覲企 一
焔
  蟆郁骸 5覯  蠏  蟆郁骸
 Best performance : 2.19%
 蟯  
 (N @ # of Filters in 1st Conv Cell)
Result(ImageNet)
 譟危  design model 豕螻 焔(SENet)螻 狩 焔 
 焔ル 讌襷 FLOP螳 蠍一ヾ modelれ 觜
Result(ImageNet)
  貉危  蟆(ex, Embedded board, Mobile device)
蟆 覈碁 觜  焔
 N螻 1st conv cell filter 螳襯 譟一 譬 焔 覲伎
Result(COCO_detection)
 ImageNet pretrained NASNet-A Faster-RCNN 蟲譟一 
 襷谿螳讌襦 蠍一ヾ SOTA 襷れ 觜 譬 焔!
 Faster-RCNN 蟲譟一 CNN network襷  蟆郁骸  YOLO, RetinaNet 觜伎  焔
 Classification target伎襷 譟磯螳 Detection, Segmentation 譬 焔 蠍磯?!
Result(Random Search 觜蟲)
 螳旧  , Random Search襯   焔 谿 觜蟲
 RL  螳 RS襯   覲企
 NAS RL vs RS 蠏碁 觜 蟆谿螳 譴企   Search space螳  讀覈
Discussion
 CNN, Classification 蟆朱 譬 焔リ骸 レ煙 覲伎譴 郁規
 企語 size螳  蟆曙一 譬 焔 蠍磯
 譟一 一危一 譬 覿襯 焔レ  蟆朱 螳!
  蠍  螳(500GPU, 4days)  ENAS!
 Detection, Segmentation 讌 蟯覈 襷 郁規 れ 
 Classification 覲牛朱 譟磯螳  蟆朱 蠍磯

More Related Content

What's hot (20)

PR-203: Class-Balanced Loss Based on Effective Number of Samples
PR-203: Class-Balanced Loss Based on Effective Number of SamplesPR-203: Class-Balanced Loss Based on Effective Number of Samples
PR-203: Class-Balanced Loss Based on Effective Number of Samples
Sunghoon Joo
carrier of_tricks_for_image_classification
carrier of_tricks_for_image_classificationcarrier of_tricks_for_image_classification
carrier of_tricks_for_image_classification
LEE HOSEONG
[蟲] Neural Architecture Search with Reinforcement Learning
[蟲] Neural Architecture Search with Reinforcement Learning[蟲] Neural Architecture Search with Reinforcement Learning
[蟲] Neural Architecture Search with Reinforcement Learning
Kiho Suh
CNN Architecture A to Z
CNN Architecture A to ZCNN Architecture A to Z
CNN Architecture A to Z
LEE HOSEONG
PR-339: Maintaining discrimination and fairness in class incremental learning
PR-339: Maintaining discrimination and fairness in class incremental learningPR-339: Maintaining discrimination and fairness in class incremental learning
PR-339: Maintaining discrimination and fairness in class incremental learning
Sunghoon Joo
PR-218: MFAS: Multimodal Fusion Architecture Search
PR-218: MFAS: Multimodal Fusion Architecture SearchPR-218: MFAS: Multimodal Fusion Architecture Search
PR-218: MFAS: Multimodal Fusion Architecture Search
Sunghoon Joo
AlexNet, VGG, GoogleNet, Resnet
AlexNet, VGG, GoogleNet, ResnetAlexNet, VGG, GoogleNet, Resnet
AlexNet, VGG, GoogleNet, Resnet
Jungwon Kim
Deep learning seminar_snu_161031
Deep learning seminar_snu_161031Deep learning seminar_snu_161031
Deep learning seminar_snu_161031
Jinwon Lee
Review MLP Mixer
Review MLP MixerReview MLP Mixer
Review MLP Mixer
Woojin Jeong
Encoding in Style: a Style Encoder for Image-to-Image Translation
Encoding in Style: a Style Encoder for Image-to-Image TranslationEncoding in Style: a Style Encoder for Image-to-Image Translation
Encoding in Style: a Style Encoder for Image-to-Image Translation
taeseon ryu
Imagination-Augmented Agents for Deep Reinforcement Learning
Imagination-Augmented Agents for Deep Reinforcement LearningImagination-Augmented Agents for Deep Reinforcement Learning
Imagination-Augmented Agents for Deep Reinforcement Learning
煙 豕
Yolo v2 urop 覦襭
Yolo v2 urop 覦襭Yolo v2 urop 覦襭
Yolo v2 urop 覦襭
DaeHeeKim31
Image net classification with deep convolutional neural networks
Image net classification with deep convolutional neural networks Image net classification with deep convolutional neural networks
Image net classification with deep convolutional neural networks
Korea, Sejong University.
る誤 ル
る誤  ルる誤  ル
る誤 ル
NAVER Engineering
Yolo v1 urop 覦襭
Yolo v1 urop 覦襭Yolo v1 urop 覦襭
Yolo v1 urop 覦襭
DaeHeeKim31
Lecture 4: Neural Networks I
Lecture 4: Neural Networks ILecture 4: Neural Networks I
Lecture 4: Neural Networks I
Sang Jun Lee
Deep Learning & Convolutional Neural Network
Deep Learning & Convolutional Neural NetworkDeep Learning & Convolutional Neural Network
Deep Learning & Convolutional Neural Network
agdatalab
ル 朱語所鍵 efficient netv2 朱碁Μ觀
ル 朱語所鍵 efficient netv2  朱碁Μ觀ル 朱語所鍵 efficient netv2  朱碁Μ觀
ル 朱語所鍵 efficient netv2 朱碁Μ觀
taeseon ryu
[paper review] 蠏觜 - Eye in the sky & 3D human pose estimation in video with ...
[paper review] 蠏觜 - Eye in the sky & 3D human pose estimation in video with ...[paper review] 蠏觜 - Eye in the sky & 3D human pose estimation in video with ...
[paper review] 蠏觜 - Eye in the sky & 3D human pose estimation in video with ...
Gyubin Son
Review SRGAN
Review SRGANReview SRGAN
Review SRGAN
Woojin Jeong
PR-203: Class-Balanced Loss Based on Effective Number of Samples
PR-203: Class-Balanced Loss Based on Effective Number of SamplesPR-203: Class-Balanced Loss Based on Effective Number of Samples
PR-203: Class-Balanced Loss Based on Effective Number of Samples
Sunghoon Joo
carrier of_tricks_for_image_classification
carrier of_tricks_for_image_classificationcarrier of_tricks_for_image_classification
carrier of_tricks_for_image_classification
LEE HOSEONG
[蟲] Neural Architecture Search with Reinforcement Learning
[蟲] Neural Architecture Search with Reinforcement Learning[蟲] Neural Architecture Search with Reinforcement Learning
[蟲] Neural Architecture Search with Reinforcement Learning
Kiho Suh
CNN Architecture A to Z
CNN Architecture A to ZCNN Architecture A to Z
CNN Architecture A to Z
LEE HOSEONG
PR-339: Maintaining discrimination and fairness in class incremental learning
PR-339: Maintaining discrimination and fairness in class incremental learningPR-339: Maintaining discrimination and fairness in class incremental learning
PR-339: Maintaining discrimination and fairness in class incremental learning
Sunghoon Joo
PR-218: MFAS: Multimodal Fusion Architecture Search
PR-218: MFAS: Multimodal Fusion Architecture SearchPR-218: MFAS: Multimodal Fusion Architecture Search
PR-218: MFAS: Multimodal Fusion Architecture Search
Sunghoon Joo
AlexNet, VGG, GoogleNet, Resnet
AlexNet, VGG, GoogleNet, ResnetAlexNet, VGG, GoogleNet, Resnet
AlexNet, VGG, GoogleNet, Resnet
Jungwon Kim
Deep learning seminar_snu_161031
Deep learning seminar_snu_161031Deep learning seminar_snu_161031
Deep learning seminar_snu_161031
Jinwon Lee
Review MLP Mixer
Review MLP MixerReview MLP Mixer
Review MLP Mixer
Woojin Jeong
Encoding in Style: a Style Encoder for Image-to-Image Translation
Encoding in Style: a Style Encoder for Image-to-Image TranslationEncoding in Style: a Style Encoder for Image-to-Image Translation
Encoding in Style: a Style Encoder for Image-to-Image Translation
taeseon ryu
Imagination-Augmented Agents for Deep Reinforcement Learning
Imagination-Augmented Agents for Deep Reinforcement LearningImagination-Augmented Agents for Deep Reinforcement Learning
Imagination-Augmented Agents for Deep Reinforcement Learning
煙 豕
Yolo v2 urop 覦襭
Yolo v2 urop 覦襭Yolo v2 urop 覦襭
Yolo v2 urop 覦襭
DaeHeeKim31
Image net classification with deep convolutional neural networks
Image net classification with deep convolutional neural networks Image net classification with deep convolutional neural networks
Image net classification with deep convolutional neural networks
Korea, Sejong University.
Yolo v1 urop 覦襭
Yolo v1 urop 覦襭Yolo v1 urop 覦襭
Yolo v1 urop 覦襭
DaeHeeKim31
Lecture 4: Neural Networks I
Lecture 4: Neural Networks ILecture 4: Neural Networks I
Lecture 4: Neural Networks I
Sang Jun Lee
Deep Learning & Convolutional Neural Network
Deep Learning & Convolutional Neural NetworkDeep Learning & Convolutional Neural Network
Deep Learning & Convolutional Neural Network
agdatalab
ル 朱語所鍵 efficient netv2 朱碁Μ觀
ル 朱語所鍵 efficient netv2  朱碁Μ觀ル 朱語所鍵 efficient netv2  朱碁Μ觀
ル 朱語所鍵 efficient netv2 朱碁Μ觀
taeseon ryu
[paper review] 蠏觜 - Eye in the sky & 3D human pose estimation in video with ...
[paper review] 蠏觜 - Eye in the sky & 3D human pose estimation in video with ...[paper review] 蠏觜 - Eye in the sky & 3D human pose estimation in video with ...
[paper review] 蠏觜 - Eye in the sky & 3D human pose estimation in video with ...
Gyubin Son

Similar to "Learning transferable architectures for scalable image recognition" Paper Review (20)

Automl
AutomlAutoml
Automl
AUTOML
AUTOMLAUTOML
AUTOML
Alexnet paper review
Alexnet paper reviewAlexnet paper review
Alexnet paper review
Image Deep Learning る伎
Image Deep Learning る伎Image Deep Learning る伎
Image Deep Learning る伎
Youngjae Kim
Summary in recent advances in deep learning for object detection
Summary in recent advances in deep learning for object detectionSummary in recent advances in deep learning for object detection
Summary in recent advances in deep learning for object detection
谿所鍵 覓
Summary in recent advances in deep learning for object detection
Summary in recent advances in deep learning for object detectionSummary in recent advances in deep learning for object detection
Summary in recent advances in deep learning for object detection
谿所鍵 覓
求== 求==梶求午メ求 8
求== 求==梶求午メ求 8求== 求==梶求午メ求 8
求== 求==梶求午メ求 8
Sunggon Song
[覿ろ語 Tech Talk] 覦一_Structure of Model and Task
[覿ろ語 Tech Talk] 覦一_Structure of Model and Task[覿ろ語 Tech Talk] 覦一_Structure of Model and Task
[覿ろ語 Tech Talk] 覦一_Structure of Model and Task
CONNECT FOUNDATION
Deep learning 求
Deep learning 求Deep learning 求
Deep learning 求
KwangPil Hong
Architecture Design Strategies
Architecture Design StrategiesArchitecture Design Strategies
Architecture Design Strategies
覯譴 蟾
History of Vision AI
History of Vision AIHistory of Vision AI
History of Vision AI
Tae Young Lee
Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)
Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)
Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)
jungminchung
[Paper Review] Visualizing and understanding convolutional networks
[Paper Review] Visualizing and understanding convolutional networks[Paper Review] Visualizing and understanding convolutional networks
[Paper Review] Visualizing and understanding convolutional networks
Korea, Sejong University.
Deep Object Detectors #1 (~2016.6)
Deep Object Detectors #1 (~2016.6)Deep Object Detectors #1 (~2016.6)
Deep Object Detectors #1 (~2016.6)
Ildoo Kim
20200309 (FSRI) deep-family_v2-br31_rabbit
20200309 (FSRI)  deep-family_v2-br31_rabbit20200309 (FSRI)  deep-family_v2-br31_rabbit
20200309 (FSRI) deep-family_v2-br31_rabbit
jason min
郁規 碁碁 Show and tell google image captioning
郁規 碁碁 Show and tell google image captioning郁規 碁碁 Show and tell google image captioning
郁規 碁碁 Show and tell google image captioning
hkh
Deep neural networks cnn rnn_ae_some practical techniques
Deep neural networks cnn rnn_ae_some practical techniquesDeep neural networks cnn rnn_ae_some practical techniques
Deep neural networks cnn rnn_ae_some practical techniques
Kang Pilsung
語襷り碓讌5乏
語襷り碓讌5乏語襷り碓讌5乏
語襷り碓讌5乏
kr_delcam
Convolutional rnn
Convolutional rnnConvolutional rnn
Convolutional rnn
Lee Gyeong Hoon
LeNet & GoogLeNet
LeNet & GoogLeNetLeNet & GoogLeNet
LeNet & GoogLeNet
Institute of Agricultural Machinery, NARO
Automl
AutomlAutoml
Automl
AUTOML
AUTOMLAUTOML
AUTOML
Alexnet paper review
Alexnet paper reviewAlexnet paper review
Alexnet paper review
Image Deep Learning る伎
Image Deep Learning る伎Image Deep Learning る伎
Image Deep Learning る伎
Youngjae Kim
Summary in recent advances in deep learning for object detection
Summary in recent advances in deep learning for object detectionSummary in recent advances in deep learning for object detection
Summary in recent advances in deep learning for object detection
谿所鍵 覓
Summary in recent advances in deep learning for object detection
Summary in recent advances in deep learning for object detectionSummary in recent advances in deep learning for object detection
Summary in recent advances in deep learning for object detection
谿所鍵 覓
求== 求==梶求午メ求 8
求== 求==梶求午メ求 8求== 求==梶求午メ求 8
求== 求==梶求午メ求 8
Sunggon Song
[覿ろ語 Tech Talk] 覦一_Structure of Model and Task
[覿ろ語 Tech Talk] 覦一_Structure of Model and Task[覿ろ語 Tech Talk] 覦一_Structure of Model and Task
[覿ろ語 Tech Talk] 覦一_Structure of Model and Task
CONNECT FOUNDATION
Deep learning 求
Deep learning 求Deep learning 求
Deep learning 求
KwangPil Hong
Architecture Design Strategies
Architecture Design StrategiesArchitecture Design Strategies
Architecture Design Strategies
覯譴 蟾
History of Vision AI
History of Vision AIHistory of Vision AI
History of Vision AI
Tae Young Lee
Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)
Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)
Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)
jungminchung
[Paper Review] Visualizing and understanding convolutional networks
[Paper Review] Visualizing and understanding convolutional networks[Paper Review] Visualizing and understanding convolutional networks
[Paper Review] Visualizing and understanding convolutional networks
Korea, Sejong University.
Deep Object Detectors #1 (~2016.6)
Deep Object Detectors #1 (~2016.6)Deep Object Detectors #1 (~2016.6)
Deep Object Detectors #1 (~2016.6)
Ildoo Kim
20200309 (FSRI) deep-family_v2-br31_rabbit
20200309 (FSRI)  deep-family_v2-br31_rabbit20200309 (FSRI)  deep-family_v2-br31_rabbit
20200309 (FSRI) deep-family_v2-br31_rabbit
jason min
郁規 碁碁 Show and tell google image captioning
郁規 碁碁 Show and tell google image captioning郁規 碁碁 Show and tell google image captioning
郁規 碁碁 Show and tell google image captioning
hkh
Deep neural networks cnn rnn_ae_some practical techniques
Deep neural networks cnn rnn_ae_some practical techniquesDeep neural networks cnn rnn_ae_some practical techniques
Deep neural networks cnn rnn_ae_some practical techniques
Kang Pilsung
語襷り碓讌5乏
語襷り碓讌5乏語襷り碓讌5乏
語襷り碓讌5乏
kr_delcam

More from LEE HOSEONG (14)

Unsupervised anomaly detection using style distillation
Unsupervised anomaly detection using style distillationUnsupervised anomaly detection using style distillation
Unsupervised anomaly detection using style distillation
LEE HOSEONG
do adversarially robust image net models transfer better
do adversarially robust image net models transfer betterdo adversarially robust image net models transfer better
do adversarially robust image net models transfer better
LEE HOSEONG
"The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Gen...
"The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Gen..."The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Gen...
"The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Gen...
LEE HOSEONG
Mixed Precision Training Review
Mixed Precision Training ReviewMixed Precision Training Review
Mixed Precision Training Review
LEE HOSEONG
MVTec AD: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection
MVTec AD: A Comprehensive Real-World Dataset for Unsupervised Anomaly DetectionMVTec AD: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection
MVTec AD: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection
LEE HOSEONG
YOLOv4: optimal speed and accuracy of object detection review
YOLOv4: optimal speed and accuracy of object detection reviewYOLOv4: optimal speed and accuracy of object detection review
YOLOv4: optimal speed and accuracy of object detection review
LEE HOSEONG
FixMatch:simplifying semi supervised learning with consistency and confidence
FixMatch:simplifying semi supervised learning with consistency and confidenceFixMatch:simplifying semi supervised learning with consistency and confidence
FixMatch:simplifying semi supervised learning with consistency and confidence
LEE HOSEONG
"Revisiting self supervised visual representation learning" Paper Review
"Revisiting self supervised visual representation learning" Paper Review"Revisiting self supervised visual representation learning" Paper Review
"Revisiting self supervised visual representation learning" Paper Review
LEE HOSEONG
Unsupervised visual representation learning overview: Toward Self-Supervision
Unsupervised visual representation learning overview: Toward Self-SupervisionUnsupervised visual representation learning overview: Toward Self-Supervision
Unsupervised visual representation learning overview: Toward Self-Supervision
LEE HOSEONG
Human uncertainty makes classification more robust, ICCV 2019 Review
Human uncertainty makes classification more robust, ICCV 2019 ReviewHuman uncertainty makes classification more robust, ICCV 2019 Review
Human uncertainty makes classification more robust, ICCV 2019 Review
LEE HOSEONG
Single Image Super Resolution Overview
Single Image Super Resolution OverviewSingle Image Super Resolution Overview
Single Image Super Resolution Overview
LEE HOSEONG
2019 ICLR Best Paper Review
2019 ICLR Best Paper Review2019 ICLR Best Paper Review
2019 ICLR Best Paper Review
LEE HOSEONG
2019 cvpr paper_overview
2019 cvpr paper_overview2019 cvpr paper_overview
2019 cvpr paper_overview
LEE HOSEONG
Pelee: a real time object detection system on mobile devices Paper Review
Pelee: a real time object detection system on mobile devices Paper ReviewPelee: a real time object detection system on mobile devices Paper Review
Pelee: a real time object detection system on mobile devices Paper Review
LEE HOSEONG
Unsupervised anomaly detection using style distillation
Unsupervised anomaly detection using style distillationUnsupervised anomaly detection using style distillation
Unsupervised anomaly detection using style distillation
LEE HOSEONG
do adversarially robust image net models transfer better
do adversarially robust image net models transfer betterdo adversarially robust image net models transfer better
do adversarially robust image net models transfer better
LEE HOSEONG
"The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Gen...
"The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Gen..."The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Gen...
"The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Gen...
LEE HOSEONG
Mixed Precision Training Review
Mixed Precision Training ReviewMixed Precision Training Review
Mixed Precision Training Review
LEE HOSEONG
MVTec AD: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection
MVTec AD: A Comprehensive Real-World Dataset for Unsupervised Anomaly DetectionMVTec AD: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection
MVTec AD: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection
LEE HOSEONG
YOLOv4: optimal speed and accuracy of object detection review
YOLOv4: optimal speed and accuracy of object detection reviewYOLOv4: optimal speed and accuracy of object detection review
YOLOv4: optimal speed and accuracy of object detection review
LEE HOSEONG
FixMatch:simplifying semi supervised learning with consistency and confidence
FixMatch:simplifying semi supervised learning with consistency and confidenceFixMatch:simplifying semi supervised learning with consistency and confidence
FixMatch:simplifying semi supervised learning with consistency and confidence
LEE HOSEONG
"Revisiting self supervised visual representation learning" Paper Review
"Revisiting self supervised visual representation learning" Paper Review"Revisiting self supervised visual representation learning" Paper Review
"Revisiting self supervised visual representation learning" Paper Review
LEE HOSEONG
Unsupervised visual representation learning overview: Toward Self-Supervision
Unsupervised visual representation learning overview: Toward Self-SupervisionUnsupervised visual representation learning overview: Toward Self-Supervision
Unsupervised visual representation learning overview: Toward Self-Supervision
LEE HOSEONG
Human uncertainty makes classification more robust, ICCV 2019 Review
Human uncertainty makes classification more robust, ICCV 2019 ReviewHuman uncertainty makes classification more robust, ICCV 2019 Review
Human uncertainty makes classification more robust, ICCV 2019 Review
LEE HOSEONG
Single Image Super Resolution Overview
Single Image Super Resolution OverviewSingle Image Super Resolution Overview
Single Image Super Resolution Overview
LEE HOSEONG
2019 ICLR Best Paper Review
2019 ICLR Best Paper Review2019 ICLR Best Paper Review
2019 ICLR Best Paper Review
LEE HOSEONG
2019 cvpr paper_overview
2019 cvpr paper_overview2019 cvpr paper_overview
2019 cvpr paper_overview
LEE HOSEONG
Pelee: a real time object detection system on mobile devices Paper Review
Pelee: a real time object detection system on mobile devices Paper ReviewPelee: a real time object detection system on mobile devices Paper Review
Pelee: a real time object detection system on mobile devices Paper Review
LEE HOSEONG

"Learning transferable architectures for scalable image recognition" Paper Review

  • 1. Learning Transferable Architectures for Scalable Image Recognition Barret Zoph, Vijay Vasudevanm Jonathon Shlens, Quoc V. Le Google Brain
  • 3. Introduction Neural Architecture Search with Reinforcement Learning(NAS) 朱 NAS : CNN(CIFAR-10), RNN(Penn Treebank) 蟲譟 り CIFAR-10 企朱 一危一 旧 800 GPU, 28days 一危一?? Learning Transferable Architectures for Scalable Image Recognition (2018, CVPR) CNN 豐 襷豢 伎 螳ロ螻 architecture襯 (NASNet) CIFAR-10朱 谿場 Convolution Cell 伎 ImageNet SOTA 焔 ! NAS 觜 旧 螳 豢 (500GPU, 4days , x7 speed up) * NAS : Nvidia K40 GPU / NASNet : Nvidia P100s GPU
  • 4. Contribution(1) Search Space ろ語 豌企ゼ 覯 り(NAS) Normal Cell, Reduction Cell 2螳 Convolution Cell 牛 り Block Cell Architecture Block
  • 5. Contribution(2) Transferability & Good Performance CIFAR-10朱 谿場 Convolution Cell 伎 Dataset 螳 CIFAR-10, ImageNet 蠍一 SOTA 焔 Object Detection NASNet 蟲譟磯ゼ 焔
  • 6. Method(1) Overall architectures of the convolutional nets are manually predetermined Normal Cell - convolutional cells that return a feature map of the same dimension Reduction Cell - convolutional cells that return a feature map where the feature map height and width is reduced by a factor of two (Initial operation with stride of two conv) Using common heuristic to double the number of filters in the output whenever the spatial activation size is reduced N, 豌 convolution cell filter 螳 User螳 讌伎殊伎 覲
  • 7. Method(2) Overall architecture 蟲 Controller(LSTM)襯 伎 Normal Cell, Reduction Cell 1 Block = hidden state input 2螳 + operation 2螳 + combine(add or concat) 1螳 豐 5螳 覓語 1螳 Cell B螳 Block朱 蟲 (朱語 B=5 焔 一) Normal Cell 企 predictions 5B螳 + Reduction Cell 企 predictions 5B螳 Block
  • 8. Method(2) Overall architecture 蟲 (Cont.) Controller(LSTM)襯 伎 Normal Cell, Reduction Cell 1 Block = hidden state input 2螳 + operation 2螳 + combine(add or concat) 1螳 豐 5螳 覓語 Operation 譬襯 れ 覦覯襦 譴 13螳 stride = 1 (Normal Cell) / stride = 1 or 2 (Reduction Cell)
  • 9. Method(2) - Supplement Overall architecture 蟲(Cont.) stride = 1 or 2 (Reduction Cell) (豢豌 : https://github.com/titu1994/Keras-NASNet/blob/master/nasnet.py)
  • 10. Method(2) NASNet-A 揃 揃 揃 <Normal Cell> 揃 揃 揃 <Reduction Cell> B螳 B螳
  • 11. Training with RL Controller(LSTM)襯 伎 豸″ Cell 牛 Architecture 螳 螻襴讀朱 REINFORCE Proximal Policy Optimization(PPO) PPO : 2017, OpenAI 豌伎 覦覯 NAS 蟇一 State : controller hidden state Action : controller螳 燕 predictions Reward : Overall architecture accuracy
  • 12. Result(CIFAR-10) 蠍一ヾ SOTA襦 れ network 觜 一 焔! NAS CNN architecture覲企 一 焔 蟆郁骸 5覯 蠏 蟆郁骸 Best performance : 2.19% 蟯 (N @ # of Filters in 1st Conv Cell)
  • 13. Result(ImageNet) 譟危 design model 豕螻 焔(SENet)螻 狩 焔 焔ル 讌襷 FLOP螳 蠍一ヾ modelれ 觜
  • 14. Result(ImageNet) 貉危 蟆(ex, Embedded board, Mobile device) 蟆 覈碁 觜 焔 N螻 1st conv cell filter 螳襯 譟一 譬 焔 覲伎
  • 15. Result(COCO_detection) ImageNet pretrained NASNet-A Faster-RCNN 蟲譟一 襷谿螳讌襦 蠍一ヾ SOTA 襷れ 觜 譬 焔! Faster-RCNN 蟲譟一 CNN network襷 蟆郁骸 YOLO, RetinaNet 觜伎 焔 Classification target伎襷 譟磯螳 Detection, Segmentation 譬 焔 蠍磯?!
  • 16. Result(Random Search 觜蟲) 螳旧 , Random Search襯 焔 谿 觜蟲 RL 螳 RS襯 覲企 NAS RL vs RS 蠏碁 觜 蟆谿螳 譴企 Search space螳 讀覈
  • 17. Discussion CNN, Classification 蟆朱 譬 焔リ骸 レ煙 覲伎譴 郁規 企語 size螳 蟆曙一 譬 焔 蠍磯 譟一 一危一 譬 覿襯 焔レ 蟆朱 螳! 蠍 螳(500GPU, 4days) ENAS! Detection, Segmentation 讌 蟯覈 襷 郁規 れ Classification 覲牛朱 譟磯螳 蟆朱 蠍磯