The document discusses geometric deep learning and spectral graph theory. It introduces convolutional neural networks (CNNs) and explains that they work well because image data has a grid structure that allows for weight sharing and translation invariance/equivariance. CNNs use filters with compact support on grid data and allow subsampling through pooling. The document then discusses spectral graph theory and spectral CNNs, introducing concepts like the graph Laplacian, eigenvectors of the Laplacian representing Fourier vectors, and spectral CNNs performing convolutions in the spectral domain using eigenvectors of the graph Laplacian.
2. Convolutional Theorem
DMIS Presentation 2
Why did Convolutional Neural Networks work so well?
Coordinates of data representation has grid
structure
Convolutional Neural Networks
Translational Invariance/Equivariance
3. Convolutional Theorem
DMIS Presentation 3
Why did Convolutional Neural Networks work so well?
Coordinates of data representation has grid
structure
Convolutional Neural Networks
Translational Invariance/Equivariance
Grids in 2D image formats : pixels
4. Convolutional Theorem
DMIS Presentation 4
Why did Convolutional Neural Networks work so well?
Coordinates of data representation has grid
structure
Convolutional Neural Networks
Translational Invariance/Equivariance
같은 물체는 translation 이후에도 같은 feature
vector로 맵핑이 이루어져야 한다.
Invariance를 강화하기 위한 기법으로 가장
대표적인 것이 Image Augmentation.
회전된 feature vector가 같아지게끔 학습한다.
5. Convolutional Theorem
DMIS Presentation 5
Why did Convolutional Neural Networks work so well?
Coordinates of data representation has grid
structure
Convolutional Neural Networks
Translational Invariance/Equivariance
같은 물체에 대해 translation 한 결과는 그 feature vector에
translation을 한 결과와 같다.
Equivariance를 강화하기 위한 기법으로 가장
대표적인 것이 Group Convnet & Capsule Net
회전된 feature vector는 둘 다 ‘비행기'라고 학습한다.
6. Convolutional Theorem
DMIS Presentation 6
Why did Convolutional Neural Networks work so well?
Coordinates of data representation has grid
structure
Convolutional Neural Networks
Translational Invariance/Equivariance
Translational structure allows weight sharing
Grid based metric allows compactly supported
filters
Multiscale dyadic clustering allows subsampling
input크기와 무관하게 적은 개수의 parameter개수로 학습
stride convolutions & pooling
7. Convolutional Theorem
DMIS Presentation 7
Why did Convolutional Neural Networks work so well?
Coordinates of data representation has grid
structure
Convolutional Neural Networks
Translational Invariance/Equivariance
Translational structure allows weight sharing
Grid based metric allows compactly supported
filters
Multiscale dyadic clustering allows subsampling
input크기와 무관하게 적은 개수의 parameter개수로 학습
stride convolutions & pooling
dyadic refers to a domain with two abstract sets of
objects 𝑋𝑋 = 𝑥𝑥1,… , 𝑥𝑥𝑁𝑁 and 𝑦𝑦 = {𝑦𝑦1, … , 𝑦𝑦𝑀𝑀} in which
observations 𝑆𝑆 are made for dyads (𝑥𝑥𝑖𝑖, 𝑦𝑦𝑘𝑘).
8. Convolutional Theorem
DMIS Presentation 8
Why did Convolutional Neural Networks work so well?
Coordinates of data representation has grid
structure
Convolutional Neural Networks
Translational Invariance/Equivariance
Translational structure allows weight sharing
Grid based metric allows compactly supported
filters
Multiscale dyadic clustering allows subsampling
input크기와 무관하게 적은 개수의 parameter개수로 학습
stride convolutions & pooling
3D-MESH
SOCIAL NETWORK
GRAPH
FUNCTION ON
MESH (HEAT)
POINT CLOUD
9. Convolutional Theorem
DMIS Presentation 9
Why did Convolutional Neural Networks work so well?
Coordinates of data representation has grid
structure
Convolutional Neural Networks
Translational Invariance/Equivariance
Translational structure allows weight sharing
Grid based metric allows compactly supported
filters
Multiscale dyadic clustering allows subsampling
input크기와 무관하게 적은 개수의 parameter개수로 학습
stride convolutions & pooling
Spatial Construction
Spectral Construction
multiscale clustering
SCNN & Coarsening
10. Convolutional Theorem
DMIS Presentation 10
Why did Convolutional Neural Networks work so well?
Coordinates of data representation has grid
structure
Convolutional Neural Networks
Translational Invariance/Equivariance
Translational structure allows weight sharing
Grid based metric allows compactly supported
filters
Multiscale dyadic clustering allows subsampling
input크기와 무관하게 적은 개수의 parameter개수로 학습
stride convolutions & pooling
Spatial Construction
Spectral Construction
multiscale clustering
SCNN & Coarsening
13. Spectral Graph Theorem
DMIS Presentation 13
Simple Intuition
Riemannian manifolds
Graphs
“Eigenvectors of the graph Laplacian converge to Eigenfunction of
the Laplace-Beltrami operator on the underlying manifold”
Fourier Basis
Eigenvectors of Laplacian
Increasing magnitudes of the eigenvalues correspond
to increasing frequencies of the eigenvectors
14. Spectral Graph Theorem
DMIS Presentation 14
What is Spectral?
Spectrum of matrix representing G
Spectral Graph Theory
15. Spectral Graph Theorem
DMIS Presentation 15
What is Spectral?
Spectrum of matrix representing G
Spectral Graph Theory
Eigenvectors ordered by magnitude of corresponding eigenvalues.
Spectrum
17. Spectral Graph Theorem
DMIS Presentation 17
What is Spectral?
Spectrum of matrix representing G
Spectral Graph Theory
Eigenvectors ordered by magnitude of corresponding eigenvalues.
Spectrum
에르미트 행렬
실수 대칭 행렬
Self-adjoint matrix
𝐴𝐴 = 𝐴𝐴 ∗= ̅𝐴𝐴𝑇𝑇
Eigen values are real
Eigen vectors are orthogonal
This assures that when moving to a spectrum,
the eigen vectors will be an orthogonal basis
and its corresponding eigenvalues will be real.
18. Spectral Graph Theorem
DMIS Presentation 18
What is Spectral?
Spectrum of matrix representing G
Spectral Graph Theory
Eigenvectors ordered by magnitude of corresponding eigenvalues.
Spectrum
𝑀𝑀𝑀𝑀 = 𝜆𝜆𝜆𝜆
𝑀𝑀 − 𝜆𝜆𝜆𝜆 𝑥𝑥 = 0
det 𝑀𝑀 − 𝜆𝜆𝜆𝜆 = 0 ( 𝑥𝑥 ≠ 0)
As 𝜆𝜆 is the variable in det 𝑀𝑀 − 𝜆𝜆𝜆𝜆 = 0, det 𝑀𝑀 − 𝜆𝜆𝜆𝜆 = 0 is a 𝑛𝑛 degree polynomial.
Therefore, there are 𝑛𝑛 eigenvalues 𝜆𝜆1, 𝜆𝜆2, 𝜆𝜆3, … , 𝜆𝜆𝑛𝑛
When 𝜆𝜆1 ≤ 𝜆𝜆2 ≤ 𝜆𝜆3 ≤ … ≤ 𝜆𝜆𝑛𝑛, spectrum = {𝑥𝑥1, 𝑥𝑥2, 𝑥𝑥3, … , 𝑥𝑥𝑛𝑛}
The solution for this equation is the eigen value
∴
20. Spectral Graph Theorem
DMIS Presentation 20
What is Spectral?
Combinatorial Laplacian
Spectral Construction
Graph Laplacian
𝐿𝐿 = 𝐷𝐷 − 𝑊𝑊 𝐿𝐿 = 𝐼𝐼 − 𝐷𝐷−
1
2 𝑊𝑊𝐷𝐷−
1
2
SCNN on d-dimensional grid
On grid space, frequency and smoothness relative to W are interrelated through Laplacian.
Smoothness
If is concentrated in the low-end of the spectrum, the corresponding spatial
kernel function is smooth; conversely, if the corresponding spatial
functions is localized, m is smooth.
Therefore, to obtain a smoother kernel function, we constrain the bandwidth
of m, enabling us to learn a smaller number of parameters; varying the
smoothness of m would control the kernel size.
21. Spectral Graph Theorem
DMIS Presentation 21
What is Spectral?
Combinatorial Laplacian
Spectral Construction
Graph Laplacian
𝐿𝐿 = 𝐷𝐷 − 𝑊𝑊 𝐿𝐿 = 𝐼𝐼 − 𝐷𝐷−
1
2 𝑊𝑊𝐷𝐷−
1
2
SCNN on d-dimensional grid
On grid space, frequency and smoothness relative to W are interrelated through Laplacian.
Smoothness
22. Spectral Graph Theorem
DMIS Presentation 22
What is Spectral?
Combinatorial Laplacian
Spectral Construction
Graph Laplacian
𝐿𝐿 = 𝐷𝐷 − 𝑊𝑊 𝐿𝐿 = 𝐼𝐼 − 𝐷𝐷−
1
2 𝑊𝑊𝐷𝐷−
1
2
SCNN on d-dimensional grid
On grid space, frequency and smoothness relative to W are interrelated through Laplacian.
Smoothness
Eigenvectors of the Laplacian = Fourier vectors
Eigenvector of Combinatorial Laplacian L =
23. Spectral Graph Theorem
DMIS Presentation 23
What is Spectral?
Combinatorial Laplacian
Spectral Construction
Graph Laplacian
𝐿𝐿 = 𝐷𝐷 − 𝑊𝑊 𝐿𝐿 = 𝐼𝐼 − 𝐷𝐷−
1
2 𝑊𝑊𝐷𝐷−
1
2
SCNN on d-dimensional grid
On grid space, frequency and smoothness relative to W are interrelated through Laplacian.
Smoothness
Eigenvectors of the Laplacian = Fourier vectors
Eigenvector of Combinatorial Laplacian L =
Eigenvalues of the Laplacian = Fourier coefficients of a signal = smoothness
29. Spectral Graph Theorem
DMIS Presentation 29
Spectral Convolution
[𝑓𝑓1, 𝑓𝑓2, 𝑓𝑓3,… 𝑓𝑓𝑝𝑝] [𝑔𝑔1, 𝑔𝑔2, 𝑔𝑔3,… 𝑔𝑔𝑞𝑞]
(n x p) (n x q)(n x k)(k x k)(k x n)
30. Spectral Graph Theorem
DMIS Presentation 30
Spectral Convolution
[𝑓𝑓1, 𝑓𝑓2, 𝑓𝑓3,… 𝑓𝑓𝑝𝑝] [𝑔𝑔1, 𝑔𝑔2, 𝑔𝑔3,… 𝑔𝑔𝑞𝑞]
(n x p) (n x q)(n x k)(k x k)(k x n)
𝐽𝐽𝐽𝐽𝐽𝐽𝐽𝐽𝐽𝐽𝐽𝐽 𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓, uniquely factored as the
product of an orthogonal matrix and a symmetric
positive definite matrix
complexity O n2
→ O(n)
31. Spectral Graph Theorem
DMIS Presentation 31
Spectral Convolution
[𝑓𝑓1, 𝑓𝑓2, 𝑓𝑓3,… 𝑓𝑓𝑝𝑝] [𝑔𝑔1, 𝑔𝑔2, 𝑔𝑔3,… 𝑔𝑔𝑞𝑞]
(n x p) (n x q)(n x k)(k x k)(k x n)
(Φ𝑘𝑘Γ1,1Φ𝑘𝑘
𝑇𝑇
𝑓𝑓1)+
(Φ𝑘𝑘Γ1,2Φ𝑘𝑘
𝑇𝑇
𝑓𝑓2)+
(Φ𝑘𝑘Γ1,3Φ𝑘𝑘
𝑇𝑇
𝑓𝑓3)+
…
+(Φ𝑘𝑘Γ1,𝑝𝑝Φ𝑘𝑘
𝑇𝑇
𝑓𝑓𝑝𝑝)
𝑠𝑠1 =
𝑔𝑔1 = 𝜉𝜉( 𝑔𝑔1)
32. Spectral Graph Theorem
DMIS Presentation 32
Spectral Convolution
[𝑓𝑓1, 𝑓𝑓2, 𝑓𝑓3,… 𝑓𝑓𝑝𝑝] [𝑔𝑔1, 𝑔𝑔2, 𝑔𝑔3,… 𝑔𝑔𝑞𝑞]
(n x p) (n x q)(n x k)(k x k)(k x n)
(Φ𝑘𝑘Γ2,1Φ𝑘𝑘
𝑇𝑇
𝑓𝑓1)+
(Φ𝑘𝑘Γ2,2Φ𝑘𝑘
𝑇𝑇
𝑓𝑓2)+
(Φ𝑘𝑘Γ2,3Φ𝑘𝑘
𝑇𝑇
𝑓𝑓3)+
…
+(Φ𝑘𝑘Γ2,𝑝𝑝Φ𝑘𝑘
𝑇𝑇
𝑓𝑓𝑝𝑝)
𝑠𝑠2 =
𝑔𝑔2 = 𝜉𝜉( 𝑠𝑠2)
33. Spectral Graph Theorem
DMIS Presentation 33
Spectral Convolution
[𝑓𝑓1, 𝑓𝑓2, 𝑓𝑓3,… 𝑓𝑓𝑝𝑝] [𝑔𝑔1, 𝑔𝑔2, 𝑔𝑔3,… 𝑔𝑔𝑞𝑞]
(n x p) (n x q)(n x k)(k x k)(k x n)
(Φ𝑘𝑘Γ3,1Φ𝑘𝑘
𝑇𝑇
𝑓𝑓1)+
(Φ𝑘𝑘Γ3,2Φ𝑘𝑘
𝑇𝑇
𝑓𝑓2)+
(Φ𝑘𝑘Γ3,3Φ𝑘𝑘
𝑇𝑇
𝑓𝑓3)+
…
+(Φ𝑘𝑘Γ3,𝑝𝑝Φ𝑘𝑘
𝑇𝑇
𝑓𝑓𝑝𝑝)
𝑠𝑠3 =
𝑔𝑔3 = 𝜉𝜉( 𝑠𝑠3)
34. Spectral Graph Theorem
DMIS Presentation 34
Spectral Convolution
[𝑓𝑓1, 𝑓𝑓2, 𝑓𝑓3,… 𝑓𝑓𝑝𝑝] [𝑔𝑔1, 𝑔𝑔2, 𝑔𝑔3,… 𝑔𝑔𝑞𝑞]
(n x p) (n x q)(n x k)(k x k)(k x n)
(Φ𝑘𝑘Γ𝑞𝑞,1Φ𝑘𝑘
𝑇𝑇
𝑓𝑓1)+
(Φ𝑘𝑘Γ𝑞𝑞,2Φ𝑘𝑘
𝑇𝑇
𝑓𝑓2)+
(Φ𝑘𝑘Γ𝑞𝑞,3Φ𝑘𝑘
𝑇𝑇
𝑓𝑓3)+
…
+(Φ𝑘𝑘Γ𝑞𝑞,𝑝𝑝Φ𝑘𝑘
𝑇𝑇
𝑓𝑓𝑝𝑝)
𝑠𝑠𝑞𝑞 =
𝑔𝑔𝑞𝑞 = 𝜉𝜉( 𝑠𝑠𝑞𝑞)
36. Spectral Graph Theorem
DMIS Presentation 36
Spectral Pooling
(n x k)(k x k)(k x n) If is concentrated in the low-end of the spectrum, the corresponding spatial
kernel function is smooth; conversely, if the corresponding spatial
functions is localized, m is smooth.
Therefore, to obtain a smoother kernel function, we constrain the bandwidth
of m, enabling us to learn a smaller number of parameters; varying the
smoothness of m would control the kernel size.
Graph Coarsening
Pooling & Strided Convolution
37. Spectral Graph Theorem
DMIS Presentation 37
Spectral Pooling
(n x k)(k x k)(k x n)
Graph Coarsening
Pooling & Strided Convolution
Local filters at deeper layers in the spatial construction to be
low frequency.
Group structure does not commute with the Laplacian
38. Spectral Graph Theorem
DMIS Presentation 38
Limitations of Spectral Pooling
Graph does not have a group structure
Group structure does not commute with the Laplacian
Individual high frequency eigenvectors become highly unstable.
By grouping high frequency eigenvectors in each octave, can obtain stable information
(=Wavelet construction)
39. Reference
DMIS Presentation 39
Convolution & Laplacian :
https://www.khanacademy.org/math/differential-equations/laplace-transform/convolution-integral/v/the-convolution-and-the-laplace-transform
Equivariance / Invariance :
https://www.slideshare.net/ssuser06e0c5/brief-intro-invariance-and-equivariance
Spectral Graph Theory :
http://www.cs.yale.edu/homes/spielman/561/
Introduction to Spectral Graph Theory :
https://www.youtube.com/watch?v=01AqmIU9Su4
SyncSpecCNN :
https://arxiv.org/pdf/1612.00606.pdf
Convergence of Laplacian Eigenmaps :
http://papers.nips.cc/paper/2989-convergence-of-laplacian-eigenmaps.pdf
Laplacian Mesh Processing :
https://people.eecs.berkeley.edu/~jrs/meshpapers/Sorkine.pdf
What is Fourier Transform? :
https://www.youtube.com/watch?v=spUNpyF58BY
Harmonic Analysis on Graphs and Networks:
http://www.gretsi.fr/peyresq14/documents/1-Vandergheynst.pdf