Graph Convolutional Neural Networks
강신동
smart bean forum leader
(주)지능도시 CEO
ceo@idosi.com
2019-03-07
Smart Bean forum seminar
at Naver D2 Startup Factory Lounge
1
Smart Bean forum
2
Facebook Group
http://facebook.com/groups/smartbean2
- AI, Deep Learning etc
- Industrial Market Tech Biz Cooperation
- Smart City
- IoT
- Smart Factory
- Smart City
Speaker
3
강신동 (Shin-Dong Kang)
- smart bean forum leader
- (주)지능도시 CEO
- ceo@idosi.com
- www.idosi.com
- Deep Learning Tech Solution Provider
- Deep Learning 용역 (연구,개발)
Deep Learning of Neural Networks
4
Machine Learning, Deep Learning 접근법
5
- 기존 방식
1. 전문 지식 기반
2. 수학적 모델링
3. 입력값 입력
4. 예측 결과 출력
- 딥러닝 방식
1. 입력값과 출력값의 측정 데이터 수집
2. 수집된 출력값의 오차 줄이도록 학습
3. 모델링 자동 완성
4. 미지의 입력값 입력
5. 예측 결과 출력
Data set is the source code !
AlphaGo
6
AlphaFold
7
Google DeepMind
과학, 생명공학, 의학, 의약 산업
아미노산 1차원 연결
-> 단백질 3차원 접힘 예측 분석
Graph Information
8
emotional
recognition
Convolutional Neural Networks
9
Convolutional AutoEncoder Deep Learning
Convolution Layer
11
Convolution Operation
12
Euler
13
Graph Theory History
14
- Euler
- Seven Bridges of Königsberg
Abstraction using Graph
15
Vertex, Node, Edge
16
Knowlege Graph
17
Coal Power Plant Process Graph
18
subtree pattern of graph
19
height = 2
Graph Types
20
Graph Convolution Concept
21
Graph Convolutional Neural Networks (GCN)
22
Laplacian Operator
23
Laplacian
24
Image Laplacian
25
MRI Laplacian
26
27
Adjacency Matrix (인접행렬)
28
Adjacency Matrix of Weighted Graph
29
Directed Graph Adjacency Matrix
(out-direction)
30
31
out-direction
32
추천 시스템에서 나타나는 graph 입니다.
침대구매 --(0.6)-> 이불구매 --(0.8)-> 베개구매
자연현상이나 물리 화학적 공정을
표현할 때도 나타나는 graph 입니다.
sun --(0.3)-> cloud --(0.1)-> rain
33
34
Gaussian distribution
Gradient of Scalar
35
Divergence of Vector
36
37
Convolution Operation
38
Laplacian
39
Graph Laplacian ( L=D-W )
40
(in-weight)
Graph Convolution Layer
41
Graph Convolution Propagation
42Cij : normalized constant
Graph Convolution Layer
43
44
Convolution Kernel Size
45
Graph Convolution Kernel
46
Graphlet (index 0~72)
47
G pattern
orbit index
Word Embedding
48
49
Graphlet Degree Vector (GDV)
50
51
Graph Downsampling & Graph Pooling
52
Heavy-edge Matching for Graph Coarsening
53
HEM
Sensor Noise Filtering with Graph
54
WL subtree kernel computing (1/5)
55
Weisfeiler-Lehman subtree method
(와이스페일러-리만)
WL Multiset-labeling (2/5)
56
WL Label compression (3/5)
57
WL Relabeling (4/5)
58
WL Feature Vector (5/5)
59
Deep Graph Convolution Networks
60
WL method
SortPooling
Link WL convolution and Neural Networks
for backpropagation
1D Convolution after SortPooling layer
61
Graph Convolutional Neural Networks (GCN)
62
63
Graph with Attention
New Protein Medicine
64
Molecular Graph
65
Molecular Fingerprint Representation
66
Protein Shape Element
67
alpha helix
beta sheet
68
Protein Folding
69
Protein Folding & Energy Level
70
Protein Binding Complex
71
Protein Folding Figurative
72
Graph Representation on Protein Structure
73
AlphaFold to predict protein foldings
74
Protein Medicine Search by GCN
75
DNA Sequence Convolutional NN
76
Gene Information of DNA
77
DNA Processing with Deep Learning
78
Predicting Protein-Protein interactions
79
80
Multiple Sequence Alignment (MSA)
81
Sequence Conservation to Protein Fold Hot Spot
82
Drug Repurposing using Natural Language Text
83
Disease gene prioritization with GCN
84
Proteomic and cell biological dissection
of mitochondrial networks
85
Contact
강신동
(주)지능도시 CEO
ceo@idosi.com
Deep Learning 관련 용역 (연구,개발)
86

Graph Convolutional Neural Networks