Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
[Journal club] Graph Attention Networks
Search
Sponsored
·
Your Podcast. Everywhere. Effortlessly.
Share. Educate. Inspire. Entertain. You do you. We'll handle the rest.
→
Semantic Machine Intelligence Lab., Keio Univ.
PRO
May 19, 2022
Technology
1.9k
0
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
[Journal club] Graph Attention Networks
Semantic Machine Intelligence Lab., Keio Univ.
PRO
May 19, 2022
More Decks by Semantic Machine Intelligence Lab., Keio Univ.
See All by Semantic Machine Intelligence Lab., Keio Univ.
[Journal club] Predict Before You Explore: Predictive Planning with Specialized Memory for Embodied Question Answering
keio_smilab
PRO
0
85
[Journal club] PHyCLIP: ðð-Product of Hyperbolic Factors Unifies Hierarchy and Compositionality in Vision-Language Representation Learning
keio_smilab
PRO
0
86
[Journal club] ReMEmbR: Building and Reasoning Over Long-Horizon Spatio-Temporal Memory for Robot Navigation
keio_smilab
PRO
0
110
[Journal club] ReLaGS: Relational Language Gaussian Splatting
keio_smilab
PRO
0
120
[Journal club] Flow as the Cross-Domain Manipulation Interface
keio_smilab
PRO
0
98
Mobi-ð: Mobilizing Your Robot Learning Policy
keio_smilab
PRO
0
160
A Gentle Introduction to Transformers
keio_smilab
PRO
16
7k
FlowAR: Scale-wise Autoregressive Image Generation Meets Flow Matching
keio_smilab
PRO
0
60
[Journal club] VLA-Adapter: An Effective Paradigm for Tiny-Scale Vision-Language-Action Model
keio_smilab
PRO
1
150
Other Decks in Technology
See All in Technology
AIãåœããåã®çµç¹ã§ ãšã³ãžãã¢ã¯ã©ãè²ã€ã
nishihira
1
970
ã¯ã©ãŠãã䜿ãåŽãããäœãåŽãž / 倧å祥寺.pm 2026åå€ç¥
fujiwara3
5
1.2k
Vãã€ã³ãåæåºç€ã«ãããããŒã¿ã¢ããªã³ã°20幎å²
taromatsui_cccmkhd
4
720
VPCã»ãã¥ãªãã£å¯Ÿå¿ã®ææ°äºæ
nagisa53
1
280
çæ AI æä»£ã«ããŸäžåºŠãåãåãããã«ã€ããŠèããŠã¿ã
kazzpapa3
1
130
ããã·ã¥ããŒã"éçº"ã«ã€ã㊠ã䜿ãããããã·ã¥ããŒãã®ã€ããããã
kimichan
0
200
PHPã§äœã£ãŠåŠã¶ãªã¢ã«ã¿ã€ã é³å£°å¯Ÿè©±AIãšWebSocketå ¥é by ã ãã«ã¿
munakata
0
130
ããã¯ã«ã³ãã¡ã¬ã³ã¹äžå€§ã¹ããŒã¯ãã«ããŒã®æå人é¡åŠ â éããèªãåãé¢ä¿æ§äœã
bash0c7
1
170
Aurora MySQL 8.4ãªãªãŒã¹ïŒ Rubyistãåããããš / what-rubyist-should-prepare-for-aurora-mysql-8-4
fkmy
0
690
AIæä»£ã®éçºçç£æ§ã¯ãå人æããããŒã èšèšãž
moongift
PRO
4
2.5k
ãã¢ãã« + ããŒãã¹ãã§èªã¿è§£ã AIãšãŒãžã§ã³ãå ¥é
oracle4engineer
PRO
2
160
AIæä»£ã®Playwright掻çš(ã·ã¹ãã ãã¹ããèªååãã ㌠å®è¡ãšã³ãžã³ã«Pla ywrightãéžãã çç±)
ynisqa1988
2
940
Featured
See All Featured
The Spectacular Lies of Maps
axbom
PRO
1
870
Building Applications with DynamoDB
mza
96
7.1k
Reality Check: Gamification 10 Years Later
codingconduct
0
2.2k
30 Presentation Tips
portentint
PRO
1
350
A brief & incomplete history of âšUX Design for the World Wide Web: 1989â2019
jct
2
420
AI Search: Implications for SEO and How to Move Forward - #ShenzhenSEOConference
aleyda
1
1.3k
Connecting the Dots Between Site Speed, User Experience & Your Business [WebExpo 2025]
tammyeverts
11
970
Exploring the Power of Turbo Streams & Action Cable | RailsConf2023
kevinliebholz
37
6.5k
Docker and Python
trallard
47
4k
The Illustrated Children's Guide to Kubernetes
chrisshort
51
53k
Why Our Code Smells
bkeepers
PRO
340
58k
Imperfection Machines: The Place of Print at Facebook
scottboms
270
14k
Transcript
Petar VeliÄkoviÄ(University of Cambridge), Guillem Cucurull(Centre de Visio per Computador),
Arantxa Casanova(Centre de Visio per Computador), Adriana Romero(Montreal Institute for Learning Algorithms), Pietro Liò(University of Cambridge), Yoshua Bengio(Montreal Institute for Learning Algorithms) Graph Attention Networks VeliÄkoviÄ, Petar, et al. "Graph attention networks." ICLR 2018. æ ¶æçŸ©å¡Ÿå€§åŠ ææµŠåæç 究宀 çäžé§¿å¹³
3 ⢠GNN ã«ãã㊠Edge ã®æ å ±ã Attention ã®éã¿ãšããŠè¡šçŸãããŒããæŽæ°ããææ³ Graph Attention
Network ( GAT ) ã®ææ¡ â 䞊åååŠçãå¯èœãšãªãïŒEdge ãå«ãèšç®é床ã®äœäžã鲿¢ â Node éã®éèŠåºŠãå€ããããããšã«ããè§£éæ§ã®åäžãæåŸ â æ¢åææ³ãããé«ãã¹ã³ã¢ãç²åŸ æŠèŠ 1. GNN ãšã¯ã»GNNã®æŽå² 2. GNN ã®åŠçå 容ã«ã€ã㊠(Message Passing Neural Network ) 3. GAT ã®æ°èŠæ§
4 ⢠CNN 㯠ç»ååé¡ã»Semantic Segmentationã»æ©æ¢°ç¿»èš³ãªã©ã®ã¿ã¹ã¯ã«å¹ åºãé©å¿ â ãããã®ããŒã¿ã¯ Grid æ§é ãšããŠã®ããŒã¿è¡šçŸ
⢠Grid æ§é ãšããŠæ±ãããšãã§ããªãããŒã¿ â äŸïŒäº€éç¶²ã»Social Network â Nodeã»Edge æ°ã¯äžå®ã§ã¯ãªã â ã°ã©ããšããŠããŒã¿è¡šçŸãç²åŸããå¿ èŠ â GNN ã®ç»å Ž èæ¯ïŒGrid æ§é ã§ã¯æ±ããªãåé¡ãã°ã©ããšããŠæ±ã
5 Graph ML ã®ã¿ã¹ã¯ãšããŠäž»ã«3〠Graph focus Node focus Edge focus
ååç©ã®å®å šäºæž¬ æå±ããŒã ã®äºæž¬ ãªã³ã¯äºæž¬
6 GNN ã®æŽå²ïŒ2005幎ã«ç»å Žã»2018幎ã¯å€§ããçºå± [1] Gori+, 2005 [2] Li+, 2016 [3]
Bruna+, 2014 [4] Defferrard+, 2016 [5] Kipf+, 2017 [6] Gilmer+, 2017 [7] Monti+, 2017 [8] Hamilton+, 2017 Graph Neural Network [1] 2005 2014 Spectral network [3] 2016 ChebNets [4] Graph Convolution Network [5] 2017 2016 Gated Graph Neural Network [2] Neural Message Passing [6] 2017 MoNet [7] 2018 Graph Attention Network Graph SAGE [8] Spectral Spatial
7 GNN ã®æŽå²ïŒGNN ã®ç»å Ž ( 2005幎 ) [1] Gori+, 2005
[2] Li+, 2016 [3] Bruna+, 2014 [4] Defferrard+, 2016 [5] Kipf+, 2017 [6] Gilmer+, 2017 [7] Monti+, 2017 [8] Hamilton+, 2017 Graph Neural Network [1] 2005 2014 Spectral network [3] 2016 ChebNets [4] Graph Convolution Network [5] 2017 2016 Gated Graph Neural Network [2] Neural Message Passing [6] 2017 MoNet [7] 2018 Graph Attention Network Graph SAGE [8] Spectral Spatial Graph Neural Network ( Gori+ 2005 ) ⢠åããŒãã®æŽæ°ã«ã¯èªèº«ãšé£æ¥ããŒããäœ¿çš â¢ ã°ã©ãããã¥ãŒã©ã«ãããã¯ãŒã¯ãšããŠæ±ã
8 GNN ã®æŽå²ïŒGraph ã« CNN ãå°å ¥ ( 2017幎 ) [1]
Gori+, 2005 [2] Li+, 2016 [3] Bruna+, 2014 [4] Defferrard+, 2016 [5] Kipf+, 2017 [6] Gilmer+, 2017 [7] Monti+, 2017 [8] Hamilton+, 2017 Graph Neural Network [1] 2005 2014 Spectral network [3] 2016 ChebNets [4] Graph Convolution Network [5] 2017 2016 Gated Graph Neural Network [2] Neural Message Passing [6] 2017 MoNet [7] 2018 Graph Attention Network Graph SAGE [8] Spectral Spatial Graph Convolution Network [ Kipf+, 2017 ]
9 GNN ã®æŽå²ïŒMessage Passing ãã¬ãŒã ã¯ãŒã¯ ( 2017幎 ) [1] Gori+,
2005 [2] Li+, 2016 [3] Bruna+, 2014 [4] Defferrard+, 2016 [5] Kipf+, 2017 [6] Gilmer+, 2017 [7] Monti+, 2017 [8] Hamilton+, 2017 Graph Neural Network [1] 2005 2014 Spectral network [3] 2016 ChebNets [4] Graph Convolution Network [5] 2017 2016 Gated Graph Neural Network [2] Neural Message Passing [6] 2017 MoNet [7] 2018 Graph Attention Network Graph SAGE [8] Spectral Spatial Neural Message Passing (Gilmer+, 2017 ) ⢠Message Passing ãšãããã¬ãŒã ã¯ãŒã¯ã§ èšç®ã®é«éåãšæ±çšæ§ãäž¡ç« â MPNN â MessageïŒç¹åŸŽé倿 â PassingïŒé ç¹æŽæ° ⢠以éã® GNN ã®ãã¬ãŒã ã¯ãŒã¯ãšããŠç¢ºç«
10 MPNNã®åŠç ( 1/4 )ïŒæŽæ°ã®å šäœå ( AGGREGATEã»UPDATE ) 3 2
1 5 4 ðð ðð ðð ⊠ðð ðð ðð ⊠ðð ðð ðð ⊠ðð ðð ðð ⊠ðð ðð ðð ⊠3 2 1 5 4 ðð ðð ðð ⊠ðð ðð ðð ⊠ðð ðð ðð ⊠ðð ðð ðð ⊠ðð ðð ðð ⊠ðð ðð ðð ⊠â1 ð¡ ðð ðð ðð ⊠ðð ðð ðð ⊠ðð ðð ðð ⊠ðð ðð ðð ⊠ðŽðºðºð ðžðºðŽððž ððð·ðŽððž â2 ð¡ â3 ð¡ â4 ð¡ â1 ð¡+1
11 ⢠A set of node featuresïŒð¡ = â1 ,
â2 , ⊠, âð , âð â âð¹ ⢠A new set of node featuresïŒð¡â² = â1 â², â2 â², ⊠, âð â² , âð â âð¹â² â¢ é£æ¥è¡å㯠èªå·± Loop ãå«ããã®ãšãã MPNNã®åŠç ( 2/4 )ïŒã°ã©ãã®åæèšå® 1 1 1 1 0 1 1 1 0 0 1 1 1 0 0 1 0 0 1 1 0 0 0 1 1 â1 â5 ⊠Features per node ð¡ 飿¥è¡å (adjacency matrix ) [5,4] [5,5] 3 2 1 5 4
12 ⢠åŠç¿å¯èœãªéã¿è¡å ( ð€ðððâð¡ ððð¡ððð¥ ) ð â âð¹Ãð¹â²
⢠éã¿è¡å ð ãâð ãšæãåããã (âð â = ðâð ) MPNNã®åŠç ( 3/4 )ïŒéã¿è¡åãšæãåããã Features per node ð¡â [5,6] â1 â â5 â ⊠â1 â5 ⊠Features per node ð¡ [5,4]
13 â¢ é£æ¥è¡åãããšãªããã Node ã®ç¹åŸŽé âð â = ðâð ãå ç®ããŠ
node ãæŽæ°ãã ⢠ãã¹ãŠã® Node ã«å¯ŸããŠåæ§ã®æŽæ°ãè¡ãïŒäžé局㮠node ç¹åŸŽéãåŸã MPNNã®åŠç ( 4/4 )ïŒããŒãã®ç¹åŸŽéã®æŽæ° ( â1 ã®å Žå) Features per node ð¡â [5,6] â1 â â5 â ⊠âð â² = Ï à· ðâð(ð) âð â = Ï à· ðâð(ð) ðâð Features per node ð¡â [5,6] â1 â² â5 â² âŠ
14 ⢠å ã»ã©ã® Node ã®æŽæ°åŒ âð â² = Ï Ï
ðâð(ð) âð â = Ï Ï ðâð(ð) ðâð 㯠Edge ã®éã¿ã 1 ⢠Attention ã®èŠé ã§ Edge ã«éã¿ãã€ã㊠Node ãæŽæ° â 䞊ååŠçãå¯èœ â Node éã®éèŠåºŠãå€ããããããšã«ããè§£éæ§ã®åäžãæåŸ GAT ã®æŠèŠïŒEdge ã«éã¿ãã€ãã
15 ððð = ð ðâð , ðâð , Attention ä¿æ°
ðïŒâð¹â² à âð¹â² â â GAT ã®ææ³ ( 1/3 )ïŒððð ã®èšç® 3 2 1 5 4 ððð ððð ððð The children play in the park.
16 ððð ãæ£èŠåïŒ Î±ðð = softmax ððð = exp(ððð) Ï
ðâð(ð) exp(ððð) â ð 㯠1局㮠Feedforward Neural Network ( éã¿ãã¯ãã« ð â â2ð¹â² )ãš LeakyReLU ãããªã αðð = exp LeakyReLU ðð ðâð ||ðâð Ï ðâð(ð) exp LeakyReLU ðð ðâð ||ðâð GAT ã®ææ³ ( 2/3 )ïŒ Î±ðð ã®èšç® ððð = ð ðâð , ðâð
17 ⢠αðð ãçšã㊠âð â² ãèšç®ãã ⢠äžåŒã Multi-Head-Attention
ã«æ¡åŒµ â ðŸ 㯠head æ° â æçµå±€ã®ã¿å¹³åããšã GAT ã®ææ³ ( 3/3 )ïŒ âð â² ã®èšç®
18 ⢠Transductive ã¿ã¹ã¯ïŒãæ¢ç¥ã ãã©ãã«ãæªç¥ã®ããŒããã«ã€ããŠã©ãã«ãäºæž¬ãã â ãæ¢ç¥ããšã¯ïŒåŠç¿ããŒã¿äžã«ããŒããããã㚠⢠Inductive ã¿ã¹ã¯ïŒãæªç¥ã®ããŒããã«å¯ŸããŠã©ãã«ãäºæž¬ãã â
ããŒãã¯ãã§ã«ã°ã©ãå ã«è¿œå ããããšããŠïŒåšèŸºããŒããåŸãããšãå¯èœ â Inductive ã¿ã¹ã¯ã®ã»ããæ±çšæ§èœãæ±ãããã å®éšïŒ3çš®é¡ã® Transductive ãš1çš®é¡ã® Inductive ã§å®éš
19 ⢠Transductive ã¿ã¹ã¯ã®è©äŸ¡å°ºåºŠïŒã¯ã©ã¹åé¡ã®æ£è§£ç ⢠Inductive ã¿ã¹ã¯ã®è©äŸ¡å°ºåºŠïŒãã¹ãŠã®ã¯ã©ã¹ã«ããã Få€ ã®å¹³å çµæïŒãããã®ã¿ã¹ã¯ã§æãé«ã粟床ãç²åŸ
ç¹ã«Inductive ã§å€§ããªå¹æ
20 ⢠GNN ã«ãã㊠Edge ã®æ å ±ã Attention ã®éã¿ãšããŠè¡šçŸãããŒããæŽæ°ããææ³ Graph Attention
Network ( GAT ) ã®ææ¡ â 䞊åååŠçãå¯èœãšãªãïŒEdge ãå«ãèšç®é床ã®äœäžã鲿¢ â Node éã®éèŠåºŠãå€ããããããšã«ããè§£éæ§ã®åäžãæåŸ â æ¢åææ³ãããé«ãã¹ã³ã¢ãç²åŸ ãŸãšã