際際滷

際際滷Share a Scribd company logo
Neural Module Network(NMN)
Neural Module
Network
蟾谿(Paul Kim)
1
Abstract
Neural Module Network
Learning to compose Neural Networks for Question Answering
VQA 企語 企語  觜螻旧 讌覓語  一危一
  螳讌 朱語 Jacob Andreas螳 VQA   ろ豌
 朱語 NMN 螳 ろ豌
1. 旧 螳ロ 企 ろ語 Layout Predictor襯 
2. 企語襷  螳ロ Visual Primitive襯 knowledge base 伎 豢襦 螳ロ蟆
2
Neural Module Network
轟
朱語  NMN(Neural Module Network) 轟 譴  蟲譟郁 旧 蟆暑
覈語 螳  覈 ろ語襦 蟲焔る .
VQA一危  蠍磯朱 螳 讌覓語  ろ語襯 螻 蠏狩
NMN覈語 ろ語 讌覓語 語 蟲譟一 磯 朱 焔
3
Neural Module Network
NMN 轟
るジ 譬襯 覈 るジ 朱 蠍磯.
Attention module(蠏碁殊 dog)
green朱 螻 label覈 blue襦

NMN 覈 module 襴曙願
蟲煙 螳ロ蠍 覓語 螳 覓語
語ろ伎る 螻一 るゼ  
NMN 覦 給  讌螻
一危一 クレ 覈碁 譴
蟆朱  讌覓語 所鍵 伎
Recurrent Nework(LSTM) 
4
Training Data Input
Training Data Input
Training data 覈 3-tuple (w, x, y)襯

W : natural-language question
X : image
Y : answer
覈語 覈 {m} 讌螻 螳螳 郁 襷り覲
theta(るジ讓 蠏碁殊 W)  string
network襦 襷ろ network layout predictor P襦
蟲焔
覈語 P(w)襯 蠍磯朱 ろ語襯 語ろ伎ろ螻 x襯
レ朱  危 企 牛 覿襯 詞企
(ex. VQA 伎 豢 覈 Classifier襦 れ)
5
Modules
Modules
覈   覈 蟲煙朱 assemble   覈  覲 蟆. 企 豕
譟壱 螳ロ vision primitive  覲 企
Moduleれ 3螳讌 basic data type 伎 operation 
A. Images
B. Unnormalized attention
C. Labels
TYPE[INSTANCE](ARG, )
A. TYPE : high-level module type(Attention, Re-Attention, )
B. INSTANCES : particular instance of model under consideration
6
Attention module
Attention Module
Attend 覈 attend[c]   企語 覈 豺襯 heatmap  unnormalized
attention 燕蠍 伎 weight vector(螳螳 C  蟲覲. Ex. cat, dog, )襦
convolution 
蠏碁殊 螻願 れ  企語  譟伎伎 螻 襾語 蟲覲蟆 
7
Re-attention module
Re-Attention Module
re-attend覈 MLP ReLU襦 蟲焔 螻  attention るジ attention朱
mapping   螳
Mapping   weight 螳螳 attend覈螻 襷谿螳讌襦 C襷 蟲覲る 蟆 蠍一
Ex) re-attend[above] above朱 伎 蠏碁殊  覦レ朱 企伎 attention

8
Combination module
Combination Module
Combine覈 2螳讌 attention  attention朱 merge 蠍磯レ 螳
Ex) Combine[and] 螳 蟆曙磯  螳 input 覈 activation  襷
蟆郁骸覓朱 activation
Ex) Combine[except] 蟆曙磯  螳讌 input 譴 豌覯讌語 input activation
region螻 覯讌 input activation inactive蟆 覲蟆曙貅 蟆郁骸覓殊 詞企
9
Classification Module
Classification Module
Classification覈 attention螻 input企語襯 螳螳 朱襖  覿襦 襷ろ 
Ex) Classify[color] color 轟 region attention 覃伎 朱襖  覿襯
襴危
10
Measurement Module
Measurement Module
Measurement覈 蟆曙磯 attention襷 伎 螳 朱襖  覿襯 mapping
覈 伎  attention unnormalized願鍵 覓語 轟 object襯 detect讌
誤蟇磯 objectれ setれ 螻壱 
11
String to networks
String to Network
Natural language讌覓語 語ろ伎ろ  蟆暑朱 覲 蟆 2螳讌 ろ 螳
A. Natural language讌覓語 Layout朱 mapping  :
: 譯殊伎 讌覓語 牛   覈 誤碁り骸 覈 螳 郁屋 讌 
B. 企蟆 襷れ伎 Layout 伎 豸 ろ語襯 assemble
12
Parsing
Parsing
Stanford Parser襦 煙  universal dependency representation 詞企
Parser  kites 螳 覲旧 kite 螳  襦 lemmatization 
危 譟伎 讌 讌覓語 wh-word 企 磯 
: 覓語レ 覩語 覿覿   symbolic form 螻
ex)
what is standing in the field? -> what(stand)
What color is the tuck -> color(truck)
Is there a circle next to a square? -> is(circle, next-to(square))
13
Layout
Layout
覈 leaf attend module, internal nodes re-attend 轟 combine module, root
node YES/NO襯 牛 QAろ measure module襦 襾語 QA 蟆曙磯
classify module襦 郁屋
Parameter 郁屋 
狩 high-level 蟲譟磯ゼ 螳讌襷 螳覲 覈れ るジ instanceれ 狩蟆 batch豌襴螳
螳ロ蠍 覓語 
ex. what color is the cat? -> classify[color](attend[cat]),
where is the truck? -> classify[where](attend[truck]))
14
Answering natural
language questions
LSTM question Encoder
A. parser襷  蟆曙 讌覓語 蠍 覓語 覓語 覩碁ゼ れ朱 覦蠑語 讌襷 旧 レ
譴   覓碁 螳 蠍磯
ex) What is flying, What are flying? -> what(fly)襦 convert.
讌襷 旧 螳螳 kites kite螳 伎 
=> question encoder 一危一 syntactic(蟲覓碁) regularities襯 覈碁蟆
れ
B. semantic(覩碁) regularities 谿  .
ex) what color is the bear?朱 讌覓語 朱 bear手 牛 蟆 襴.
green企手 豢襦 蟆 伎
=> question encode 企 譬襯 螻. 讀, semantic(覩碁) regularities襯 覈碁  
15
Answering natural
language questions
LSTM question Encoder
A. parser襷  蟆曙 讌覓語 蠍 覓語 覓語 覩碁ゼ れ朱 覦蠑語 讌襷 旧 レ
譴   覓碁 螳 蠍磯
ex) What is flying, What are flying? -> what(fly)襦 convert.
讌襷 旧 螳螳 kites kite螳 伎 
=> question encoder 一危一 syntactic(蟲覓碁) regularities襯 覈碁蟆
れ
B. semantic(覩碁) regularities 谿  .
ex) what color is the bear?朱 讌覓語 朱 bear手 牛 蟆 襴.
green企手 豢襦 蟆 伎
=> question encode 企 譬襯 螻. 讀, semantic(覩碁) regularities襯 覈碁  
豕譬覈語
Neural Module Network
 Output螻
LSTM question
Encoder襯
蟆壱
16
Answering natural
language questions
1024 hidden unit 螳 standard
single-layer LSTM 
Question modeling 蟲煙
NMN root module螻 螳 旧
 覿襯 豸″. 覈碁覿 豕譬
豸′  襯 覿 geometry
average願, text 覦 image
feature襯 伎 朱 れ
reweighted
NMN螻 sequence modeling 蟲
襯  豌 覈語 螻給朱

17
Training
Optimizer
讌覓語 牛  dynamic network structure 覓語 朱 weigh るジ 蟆覲企 譯
一危碁 蟆渚レ . 企 伎 覓語 adaptive per-weight learning rate襯 螳
螻襴讀  SGD覲企 譬 焔レ 企 蟆 誤.
蠏碁 AdaDelta襯  蠏
蠍一牛 
detect[cat]  螻 語蠍磯 螻蟇磯 豐蠍壱讌 朱 combine[and] attention
蟲讌 螻壱蠍 伎 螻 蟆 朱 蟆 蠍一牛伎 
End-to-End覦 糾骸 蟆郁骸覓朱 企 behavior襯 給.
18

More Related Content

What's hot (18)

[貉危磯螻 瑚概讌] 7. 炎咳 蟆暑 2
[貉危磯螻 瑚概讌] 7. 炎咳 蟆暑 2[貉危磯螻 瑚概讌] 7. 炎咳 蟆暑 2
[貉危磯螻 瑚概讌] 7. 炎咳 蟆暑 2
jdo
Denoising auto encoders(d a)
Denoising auto encoders(d a)Denoising auto encoders(d a)
Denoising auto encoders(d a)
Tae Young Lee
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
InfoGAN Paper Review
InfoGAN Paper ReviewInfoGAN Paper Review
InfoGAN Paper Review
瑚概 蟆暑 蟲 蟯 螳 る
瑚概 蟆暑 蟲 蟯 螳 る瑚概 蟆暑 蟲 蟯 螳 る
瑚概 蟆暑 蟲 蟯 螳 る
Woonghee Lee
Chapter 7 Regularization for deep learning - 2
Chapter 7 Regularization for deep learning - 2Chapter 7 Regularization for deep learning - 2
Chapter 7 Regularization for deep learning - 2
KyeongUkJang
CNN 豐覲伎螳 襷 豐覲伎 螳企 (VGG 所 )
CNN 豐覲伎螳 襷 豐覲伎 螳企 (VGG 所 )CNN 豐覲伎螳 襷 豐覲伎 螳企 (VGG 所 )
CNN 豐覲伎螳 襷 豐覲伎 螳企 (VGG 所 )
Lee Seungeun
Training data-efficient image transformers & distillation through attention
Training data-efficient image transformers & distillation through attentionTraining data-efficient image transformers & distillation through attention
Training data-efficient image transformers & distillation through attention
taeseon ryu
Deep learning overview
Deep learning overviewDeep learning overview
Deep learning overview
螳覩手記 螳覩手記
讖讌襷ル2 - CNN RNN 覯
讖讌襷ル2 - CNN RNN 覯讖讌襷ル2 - CNN RNN 覯
讖讌襷ル2 - CNN RNN 覯
Modulabs
朱 tensorflow 蠍 - tutorial
朱 tensorflow 蠍 - tutorial朱 tensorflow 蠍 - tutorial
朱 tensorflow 蠍 - tutorial
Lee Seungeun
[貉危磯螻 瑚概讌] 8. 炎咳 蟆暑 ろ豌 5 - Others
 [貉危磯螻 瑚概讌] 8. 炎咳 蟆暑 ろ豌 5 - Others [貉危磯螻 瑚概讌] 8. 炎咳 蟆暑 ろ豌 5 - Others
[貉危磯螻 瑚概讌] 8. 炎咳 蟆暑 ろ豌 5 - Others
jdo
[ろ襷 ろ磯 1殊姶] 襴
[ろ襷 ろ磯 1殊姶] 襴[ろ襷 ろ磯 1殊姶] 襴
[ろ襷 ろ磯 1殊姶] 襴
AnselmKim
Chapter 15 Representation learning - 1
Chapter 15 Representation learning - 1Chapter 15 Representation learning - 1
Chapter 15 Representation learning - 1
KyeongUkJang
[貉危磯螻 瑚概讌] 10. 蟆暑 牛蠍 1 - 2. 一危 豌襴
[貉危磯螻 瑚概讌] 10. 蟆暑 牛蠍  1 - 2. 一危 豌襴[貉危磯螻 瑚概讌] 10. 蟆暑 牛蠍  1 - 2. 一危 豌襴
[貉危磯螻 瑚概讌] 10. 蟆暑 牛蠍 1 - 2. 一危 豌襴
jdo
Deep Learning Into Advance - 1. Image, ConvNet
Deep Learning Into Advance - 1. Image, ConvNetDeep Learning Into Advance - 1. Image, ConvNet
Deep Learning Into Advance - 1. Image, ConvNet
Hyojun Kim
貊 れ旧朱 危危 瑚概讌
貊 れ旧朱 危危 瑚概讌貊 れ旧朱 危危 瑚概讌
貊 れ旧朱 危危 瑚概讌
ル 蠍磯蓋
ル 蠍磯蓋ル 蠍磯蓋
ル 蠍磯蓋
deepseaswjh
[貉危磯螻 瑚概讌] 7. 炎咳 蟆暑 2
[貉危磯螻 瑚概讌] 7. 炎咳 蟆暑 2[貉危磯螻 瑚概讌] 7. 炎咳 蟆暑 2
[貉危磯螻 瑚概讌] 7. 炎咳 蟆暑 2
jdo
Denoising auto encoders(d a)
Denoising auto encoders(d a)Denoising auto encoders(d a)
Denoising auto encoders(d a)
Tae Young Lee
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
InfoGAN Paper Review
InfoGAN Paper ReviewInfoGAN Paper Review
InfoGAN Paper Review
瑚概 蟆暑 蟲 蟯 螳 る
瑚概 蟆暑 蟲 蟯 螳 る瑚概 蟆暑 蟲 蟯 螳 る
瑚概 蟆暑 蟲 蟯 螳 る
Woonghee Lee
Chapter 7 Regularization for deep learning - 2
Chapter 7 Regularization for deep learning - 2Chapter 7 Regularization for deep learning - 2
Chapter 7 Regularization for deep learning - 2
KyeongUkJang
CNN 豐覲伎螳 襷 豐覲伎 螳企 (VGG 所 )
CNN 豐覲伎螳 襷 豐覲伎 螳企 (VGG 所 )CNN 豐覲伎螳 襷 豐覲伎 螳企 (VGG 所 )
CNN 豐覲伎螳 襷 豐覲伎 螳企 (VGG 所 )
Lee Seungeun
Training data-efficient image transformers & distillation through attention
Training data-efficient image transformers & distillation through attentionTraining data-efficient image transformers & distillation through attention
Training data-efficient image transformers & distillation through attention
taeseon ryu
讖讌襷ル2 - CNN RNN 覯
讖讌襷ル2 - CNN RNN 覯讖讌襷ル2 - CNN RNN 覯
讖讌襷ル2 - CNN RNN 覯
Modulabs
朱 tensorflow 蠍 - tutorial
朱 tensorflow 蠍 - tutorial朱 tensorflow 蠍 - tutorial
朱 tensorflow 蠍 - tutorial
Lee Seungeun
[貉危磯螻 瑚概讌] 8. 炎咳 蟆暑 ろ豌 5 - Others
 [貉危磯螻 瑚概讌] 8. 炎咳 蟆暑 ろ豌 5 - Others [貉危磯螻 瑚概讌] 8. 炎咳 蟆暑 ろ豌 5 - Others
[貉危磯螻 瑚概讌] 8. 炎咳 蟆暑 ろ豌 5 - Others
jdo
[ろ襷 ろ磯 1殊姶] 襴
[ろ襷 ろ磯 1殊姶] 襴[ろ襷 ろ磯 1殊姶] 襴
[ろ襷 ろ磯 1殊姶] 襴
AnselmKim
Chapter 15 Representation learning - 1
Chapter 15 Representation learning - 1Chapter 15 Representation learning - 1
Chapter 15 Representation learning - 1
KyeongUkJang
[貉危磯螻 瑚概讌] 10. 蟆暑 牛蠍 1 - 2. 一危 豌襴
[貉危磯螻 瑚概讌] 10. 蟆暑 牛蠍  1 - 2. 一危 豌襴[貉危磯螻 瑚概讌] 10. 蟆暑 牛蠍  1 - 2. 一危 豌襴
[貉危磯螻 瑚概讌] 10. 蟆暑 牛蠍 1 - 2. 一危 豌襴
jdo
Deep Learning Into Advance - 1. Image, ConvNet
Deep Learning Into Advance - 1. Image, ConvNetDeep Learning Into Advance - 1. Image, ConvNet
Deep Learning Into Advance - 1. Image, ConvNet
Hyojun Kim
貊 れ旧朱 危危 瑚概讌
貊 れ旧朱 危危 瑚概讌貊 れ旧朱 危危 瑚概讌
貊 れ旧朱 危危 瑚概讌
ル 蠍磯蓋
ル 蠍磯蓋ル 蠍磯蓋
ル 蠍磯蓋
deepseaswjh

More from Yechan(Paul) Kim (8)

螳糾骸 LV&A 蠏碁Μ螻 Navigation Agent
螳糾骸 LV&A 蠏碁Μ螻 Navigation Agent螳糾骸 LV&A 蠏碁Μ螻 Navigation Agent
螳糾骸 LV&A 蠏碁Μ螻 Navigation Agent
Yechan(Paul) Kim
Learning to Communicate to Solve Riddles with Deep Distributed Recurrent Q-Ne...
Learning to Communicate to Solve Riddles with Deep Distributed Recurrent Q-Ne...Learning to Communicate to Solve Riddles with Deep Distributed Recurrent Q-Ne...
Learning to Communicate to Solve Riddles with Deep Distributed Recurrent Q-Ne...
Yechan(Paul) Kim
Multiagent Cooperative and Competition with Deep Reinforcement Learning
Multiagent Cooperative and Competition with Deep Reinforcement LearningMultiagent Cooperative and Competition with Deep Reinforcement Learning
Multiagent Cooperative and Competition with Deep Reinforcement Learning
Yechan(Paul) Kim
2018 global ai_bootcamp_seoul_HomeNavi(Reinforcement Learning, AI)
2018 global ai_bootcamp_seoul_HomeNavi(Reinforcement Learning, AI)2018 global ai_bootcamp_seoul_HomeNavi(Reinforcement Learning, AI)
2018 global ai_bootcamp_seoul_HomeNavi(Reinforcement Learning, AI)
Yechan(Paul) Kim
3D Environment : HomeNavigation
3D Environment : HomeNavigation3D Environment : HomeNavigation
3D Environment : HomeNavigation
Yechan(Paul) Kim
Diversity is all you need(DIAYN) : Learning Skills without a Reward Function
Diversity is all you need(DIAYN) : Learning Skills without a Reward FunctionDiversity is all you need(DIAYN) : Learning Skills without a Reward Function
Diversity is all you need(DIAYN) : Learning Skills without a Reward Function
Yechan(Paul) Kim
pyconkr 2018 RL_Adventure : Rainbow(value based Reinforcement Learning)
pyconkr 2018 RL_Adventure : Rainbow(value based Reinforcement Learning)pyconkr 2018 RL_Adventure : Rainbow(value based Reinforcement Learning)
pyconkr 2018 RL_Adventure : Rainbow(value based Reinforcement Learning)
Yechan(Paul) Kim
pycon2018 "RL Adventure : DQN 覿 Rainbow DQN蟾讌"
pycon2018 "RL Adventure : DQN 覿 Rainbow DQN蟾讌"pycon2018 "RL Adventure : DQN 覿 Rainbow DQN蟾讌"
pycon2018 "RL Adventure : DQN 覿 Rainbow DQN蟾讌"
Yechan(Paul) Kim
螳糾骸 LV&A 蠏碁Μ螻 Navigation Agent
螳糾骸 LV&A 蠏碁Μ螻 Navigation Agent螳糾骸 LV&A 蠏碁Μ螻 Navigation Agent
螳糾骸 LV&A 蠏碁Μ螻 Navigation Agent
Yechan(Paul) Kim
Learning to Communicate to Solve Riddles with Deep Distributed Recurrent Q-Ne...
Learning to Communicate to Solve Riddles with Deep Distributed Recurrent Q-Ne...Learning to Communicate to Solve Riddles with Deep Distributed Recurrent Q-Ne...
Learning to Communicate to Solve Riddles with Deep Distributed Recurrent Q-Ne...
Yechan(Paul) Kim
Multiagent Cooperative and Competition with Deep Reinforcement Learning
Multiagent Cooperative and Competition with Deep Reinforcement LearningMultiagent Cooperative and Competition with Deep Reinforcement Learning
Multiagent Cooperative and Competition with Deep Reinforcement Learning
Yechan(Paul) Kim
2018 global ai_bootcamp_seoul_HomeNavi(Reinforcement Learning, AI)
2018 global ai_bootcamp_seoul_HomeNavi(Reinforcement Learning, AI)2018 global ai_bootcamp_seoul_HomeNavi(Reinforcement Learning, AI)
2018 global ai_bootcamp_seoul_HomeNavi(Reinforcement Learning, AI)
Yechan(Paul) Kim
3D Environment : HomeNavigation
3D Environment : HomeNavigation3D Environment : HomeNavigation
3D Environment : HomeNavigation
Yechan(Paul) Kim
Diversity is all you need(DIAYN) : Learning Skills without a Reward Function
Diversity is all you need(DIAYN) : Learning Skills without a Reward FunctionDiversity is all you need(DIAYN) : Learning Skills without a Reward Function
Diversity is all you need(DIAYN) : Learning Skills without a Reward Function
Yechan(Paul) Kim
pyconkr 2018 RL_Adventure : Rainbow(value based Reinforcement Learning)
pyconkr 2018 RL_Adventure : Rainbow(value based Reinforcement Learning)pyconkr 2018 RL_Adventure : Rainbow(value based Reinforcement Learning)
pyconkr 2018 RL_Adventure : Rainbow(value based Reinforcement Learning)
Yechan(Paul) Kim
pycon2018 "RL Adventure : DQN 覿 Rainbow DQN蟾讌"
pycon2018 "RL Adventure : DQN 覿 Rainbow DQN蟾讌"pycon2018 "RL Adventure : DQN 覿 Rainbow DQN蟾讌"
pycon2018 "RL Adventure : DQN 覿 Rainbow DQN蟾讌"
Yechan(Paul) Kim

Neural module Network

  • 1. Neural Module Network(NMN) Neural Module Network 蟾谿(Paul Kim) 1
  • 2. Abstract Neural Module Network Learning to compose Neural Networks for Question Answering VQA 企語 企語 觜螻旧 讌覓語 一危一 螳讌 朱語 Jacob Andreas螳 VQA ろ豌 朱語 NMN 螳 ろ豌 1. 旧 螳ロ 企 ろ語 Layout Predictor襯 2. 企語襷 螳ロ Visual Primitive襯 knowledge base 伎 豢襦 螳ロ蟆 2
  • 3. Neural Module Network 轟 朱語 NMN(Neural Module Network) 轟 譴 蟲譟郁 旧 蟆暑 覈語 螳 覈 ろ語襦 蟲焔る . VQA一危 蠍磯朱 螳 讌覓語 ろ語襯 螻 蠏狩 NMN覈語 ろ語 讌覓語 語 蟲譟一 磯 朱 焔 3
  • 4. Neural Module Network NMN 轟 るジ 譬襯 覈 るジ 朱 蠍磯. Attention module(蠏碁殊 dog) green朱 螻 label覈 blue襦 NMN 覈 module 襴曙願 蟲煙 螳ロ蠍 覓語 螳 覓語 語ろ伎る 螻一 るゼ NMN 覦 給 讌螻 一危一 クレ 覈碁 譴 蟆朱 讌覓語 所鍵 伎 Recurrent Nework(LSTM) 4
  • 5. Training Data Input Training Data Input Training data 覈 3-tuple (w, x, y)襯 W : natural-language question X : image Y : answer 覈語 覈 {m} 讌螻 螳螳 郁 襷り覲 theta(るジ讓 蠏碁殊 W) string network襦 襷ろ network layout predictor P襦 蟲焔 覈語 P(w)襯 蠍磯朱 ろ語襯 語ろ伎ろ螻 x襯 レ朱 危 企 牛 覿襯 詞企 (ex. VQA 伎 豢 覈 Classifier襦 れ) 5
  • 6. Modules Modules 覈 覈 蟲煙朱 assemble 覈 覲 蟆. 企 豕 譟壱 螳ロ vision primitive 覲 企 Moduleれ 3螳讌 basic data type 伎 operation A. Images B. Unnormalized attention C. Labels TYPE[INSTANCE](ARG, ) A. TYPE : high-level module type(Attention, Re-Attention, ) B. INSTANCES : particular instance of model under consideration 6
  • 7. Attention module Attention Module Attend 覈 attend[c] 企語 覈 豺襯 heatmap unnormalized attention 燕蠍 伎 weight vector(螳螳 C 蟲覲. Ex. cat, dog, )襦 convolution 蠏碁殊 螻願 れ 企語 譟伎伎 螻 襾語 蟲覲蟆 7
  • 8. Re-attention module Re-Attention Module re-attend覈 MLP ReLU襦 蟲焔 螻 attention るジ attention朱 mapping 螳 Mapping weight 螳螳 attend覈螻 襷谿螳讌襦 C襷 蟲覲る 蟆 蠍一 Ex) re-attend[above] above朱 伎 蠏碁殊 覦レ朱 企伎 attention 8
  • 9. Combination module Combination Module Combine覈 2螳讌 attention attention朱 merge 蠍磯レ 螳 Ex) Combine[and] 螳 蟆曙磯 螳 input 覈 activation 襷 蟆郁骸覓朱 activation Ex) Combine[except] 蟆曙磯 螳讌 input 譴 豌覯讌語 input activation region螻 覯讌 input activation inactive蟆 覲蟆曙貅 蟆郁骸覓殊 詞企 9
  • 10. Classification Module Classification Module Classification覈 attention螻 input企語襯 螳螳 朱襖 覿襦 襷ろ Ex) Classify[color] color 轟 region attention 覃伎 朱襖 覿襯 襴危 10
  • 11. Measurement Module Measurement Module Measurement覈 蟆曙磯 attention襷 伎 螳 朱襖 覿襯 mapping 覈 伎 attention unnormalized願鍵 覓語 轟 object襯 detect讌 誤蟇磯 objectれ setれ 螻壱 11
  • 12. String to networks String to Network Natural language讌覓語 語ろ伎ろ 蟆暑朱 覲 蟆 2螳讌 ろ 螳 A. Natural language讌覓語 Layout朱 mapping : : 譯殊伎 讌覓語 牛 覈 誤碁り骸 覈 螳 郁屋 讌 B. 企蟆 襷れ伎 Layout 伎 豸 ろ語襯 assemble 12
  • 13. Parsing Parsing Stanford Parser襦 煙 universal dependency representation 詞企 Parser kites 螳 覲旧 kite 螳 襦 lemmatization 危 譟伎 讌 讌覓語 wh-word 企 磯 : 覓語レ 覩語 覿覿 symbolic form 螻 ex) what is standing in the field? -> what(stand) What color is the tuck -> color(truck) Is there a circle next to a square? -> is(circle, next-to(square)) 13
  • 14. Layout Layout 覈 leaf attend module, internal nodes re-attend 轟 combine module, root node YES/NO襯 牛 QAろ measure module襦 襾語 QA 蟆曙磯 classify module襦 郁屋 Parameter 郁屋 狩 high-level 蟲譟磯ゼ 螳讌襷 螳覲 覈れ るジ instanceれ 狩蟆 batch豌襴螳 螳ロ蠍 覓語 ex. what color is the cat? -> classify[color](attend[cat]), where is the truck? -> classify[where](attend[truck])) 14
  • 15. Answering natural language questions LSTM question Encoder A. parser襷 蟆曙 讌覓語 蠍 覓語 覓語 覩碁ゼ れ朱 覦蠑語 讌襷 旧 レ 譴 覓碁 螳 蠍磯 ex) What is flying, What are flying? -> what(fly)襦 convert. 讌襷 旧 螳螳 kites kite螳 伎 => question encoder 一危一 syntactic(蟲覓碁) regularities襯 覈碁蟆 れ B. semantic(覩碁) regularities 谿 . ex) what color is the bear?朱 讌覓語 朱 bear手 牛 蟆 襴. green企手 豢襦 蟆 伎 => question encode 企 譬襯 螻. 讀, semantic(覩碁) regularities襯 覈碁 15
  • 16. Answering natural language questions LSTM question Encoder A. parser襷 蟆曙 讌覓語 蠍 覓語 覓語 覩碁ゼ れ朱 覦蠑語 讌襷 旧 レ 譴 覓碁 螳 蠍磯 ex) What is flying, What are flying? -> what(fly)襦 convert. 讌襷 旧 螳螳 kites kite螳 伎 => question encoder 一危一 syntactic(蟲覓碁) regularities襯 覈碁蟆 れ B. semantic(覩碁) regularities 谿 . ex) what color is the bear?朱 讌覓語 朱 bear手 牛 蟆 襴. green企手 豢襦 蟆 伎 => question encode 企 譬襯 螻. 讀, semantic(覩碁) regularities襯 覈碁 豕譬覈語 Neural Module Network Output螻 LSTM question Encoder襯 蟆壱 16
  • 17. Answering natural language questions 1024 hidden unit 螳 standard single-layer LSTM Question modeling 蟲煙 NMN root module螻 螳 旧 覿襯 豸″. 覈碁覿 豕譬 豸′ 襯 覿 geometry average願, text 覦 image feature襯 伎 朱 れ reweighted NMN螻 sequence modeling 蟲 襯 豌 覈語 螻給朱 17
  • 18. Training Optimizer 讌覓語 牛 dynamic network structure 覓語 朱 weigh るジ 蟆覲企 譯 一危碁 蟆渚レ . 企 伎 覓語 adaptive per-weight learning rate襯 螳 螻襴讀 SGD覲企 譬 焔レ 企 蟆 誤. 蠏碁 AdaDelta襯 蠏 蠍一牛 detect[cat] 螻 語蠍磯 螻蟇磯 豐蠍壱讌 朱 combine[and] attention 蟲讌 螻壱蠍 伎 螻 蟆 朱 蟆 蠍一牛伎 End-to-End覦 糾骸 蟆郁骸覓朱 企 behavior襯 給. 18