Manifold learning

48
A Brief Introduction to Manifold Learning Wei Yang [email protected] 06/16/2022 1 Some slides are from Geometric Methods and Manifold Learning in Machine Learning (Mikhail Belkin and Partha Niyoqi). Summer School (MLSS), Chicago 2009

Transcript of Manifold learning

Page 1: Manifold learning

05/03/2023 1

A Brief Introduction to Manifold Learning

Wei Yang

[email protected]

Some slides are from Geometric Methods and Manifold Learning in Machine Learning (Mikhail Belkin and Partha Niyoqi). Summer School (MLSS), Chicago 2009

Page 2: Manifold learning

05/03/2023 2

What is a manifold?

https://en.wikipedia.org/wiki/Manifold

Page 3: Manifold learning

05/03/2023 3

Manifolds in visual perception

Consider a simple example of image variability, the set of all facial images generated by varying the orientation of a face

• 1D manifold:– a single degree of freedom: the angle of rotation

• The dimensionality of would increase if we allow– image scaling

– illumination changing

– …

The Manifold Ways of Perception•H. Sebastian Seung and •Daniel D. LeeScience 22 December 2000

Page 4: Manifold learning

05/03/2023 4

Why Manifolds?

• Euclidean distance in the high dimensional input space may not accurately reflect the intrinsic similarity– Euclidean distance

– Geodesic distance

Linear Manifold VS. Nonlinear Manifold

Page 5: Manifold learning

05/03/2023 5

Differential Geometry

• Embedded manifolds

• Tangent space

• Geodesic

• Laplace-Beltrami operator

Page 6: Manifold learning

05/03/2023 6

Embedded manifolds

• Locally (not globally) looks like Euclidean space

𝑆  2⊂ℝ3

Page 7: Manifold learning

05/03/2023 7

Example: circle

• Charts: continuous, invertible

• Atlas: charts covered the whole circle

• Transition map 𝑥

𝑦

¿𝜙 h𝑟𝑖𝑔 𝑡 (𝑎 ,√1−𝑎2)¿√1−𝑎2

𝑎

𝜙𝑡𝑜𝑝❑ − 1 (𝑎 )

𝜙 h𝑟𝑖𝑔 𝑡 (𝜙𝑡𝑜𝑝❑ −1 (𝑎) )

http://en.wikipedia.org/wiki/Manifold

Page 8: Manifold learning

05/03/2023 8

Tangent space

• -dimensional affine subspace of

𝑇 𝑝ℳ  𝑘⊂ℝ𝑁

𝑝

Page 9: Manifold learning

05/03/2023 9

Tangent vectors and curves

• Tangent vectors <———> curves.

Geometric Methods and Manifold Learning – p. 1

𝑣

𝜙 (𝑡 )

𝑝

Page 10: Manifold learning

05/03/2023 10

Tangent vectors as derivatives

• Tangent vectors <———> Directional derivatives

Geometric Methods and Manifold Learning – p. 1

𝑣

𝜙 (𝑡 )

𝑝

Page 11: Manifold learning

05/03/2023 11

Riemannian geometry

• Norms and angles in tangent space

Geometric Methods and Manifold Learning – p. 1

𝑣

𝑤𝑝

Page 12: Manifold learning

05/03/2023 12

Length of curves and geodesics

• Can measure length using norm in tangent space.

• Geodesic — shortest curve between two points.

Geometric Methods and Manifold Learning – p. 1

𝑝

Page 13: Manifold learning

05/03/2023 13

Gradient

• Tangent vectors <———> Directional derivatives

• Gradient points in the direction of maximum change.

Geometric Methods and Manifold Learning – p. 1

¿𝛻 𝑓 ,𝑣>≡ 𝑑𝑓𝑑𝑣

𝑣

𝜙 (𝑡 )

𝑝

Page 14: Manifold learning

05/03/2023 14

Exponential map

• Geodesic:

Geometric Methods and Manifold Learning – p. 1

𝑞

𝑝

𝑟

𝑣𝑤

𝜙 (𝑡 )

Page 15: Manifold learning

05/03/2023 15

Laplace-Beltrami operator

• Orthonormal coordinate system.

Geometric Methods and Manifold Learning – p. 1

𝑝 𝑥2𝑥1

exp𝑝 :𝑇 𝑝ℳ  𝑘→ℳ  𝑘

Page 16: Manifold learning

05/03/2023 16

Linear Manifold Learning

• Principal Components Analysis

• Multidimensional Scaling

Page 17: Manifold learning

05/03/2023 17

Principal Components Analysis

• Given … with mean 0

• Find such that

• And

• is leading eigenvectors of

Page 18: Manifold learning

05/03/2023 18

Multidimensional Scaling

• MDS: exploring similarities or dissimilarities in data.

• Given data points with distance function is defined as:

• The dissimilarity matrix can be defined as:

Find … such that

Page 19: Manifold learning

05/03/2023 19

Nonlinear Manifold Learning

• ISOMAP (Tenenbaum, et al, 00)

• LLE (Roweis, Saul, 00)

• Laplacian Eigenmaps (Belkin, Niyogi, 01)

Page 20: Manifold learning

05/03/2023 20

Algorithmic framework

• Neighborhood graph common to all methods.

Page 21: Manifold learning

05/03/2023 21

Isomap: Motivation

• PCA/MDS see just the Euclidean structure

• Only geodesic distances reflect the true low-dimensional geometry of the manifold

• The question: – How to approximate geodesic distances?

Page 22: Manifold learning

05/03/2023 22

Isomap

1. Construct neighborhood graph using Euclidean distance– -Isomap: neighbors within a radius

– -Isomap: nearest neighbors

2. Compute shortest path as the approximation of geodesic distance

1.

2. For , replace all by

3. Construct -dimensional embedding using MDS

Page 23: Manifold learning

05/03/2023 23

Isomap: results

Face varying in pose and illumination Hand varying in finger extension & wrist rotation

Page 24: Manifold learning

05/03/2023 24

Isomap: estimate the intrinsic dimensionality

Page 25: Manifold learning

05/03/2023 25

Locally Linear Embedding

• Intuition: each data point and its neighbors are expected to lie on or close to a locally linear patch of the manifold.

Page 26: Manifold learning

05/03/2023 26

Locally Linear Embedding

1. Assign neighbours to each data point (-NN)

2. Reconstruct each point by a weighted linear combination of its neighbors.

3. Map each point to embedded coordinates.

S T Roweis, and L K Saul Science 2000;290:2323-2326

Page 27: Manifold learning

05/03/2023 27

Steps of locally linear embedding

• Suppose we have data points in a dimensional space.

• Step 1: Construct neighborhood graph– -NN neighborhood

– Euclidean distance or normalized dot products

Page 28: Manifold learning

05/03/2023 28

Steps of locally linear embedding

• Step 2: Compute the weights that best linearly reconstruct from its neighbors by minimizing

where

Page 29: Manifold learning

05/03/2023 29

Steps of locally linear embedding

• Step 3: Compute the low-dimensional embedding best reconstructed by by minimizing

• Note: is a sparse matrix, and -th row is barycentric coordinates (center of mass) of in the basis of its nearest neighbors.

• Similar to PCA, using lowest eigenvectors of to embed.

Page 30: Manifold learning

05/03/2023 30

LLE (Comments by Ruimao Zhang)

算法优点• LLE 算法可以学习任意维的局部线性的低维流形 .

• LLE 算法中的待定参数很少 , K 和 d.

• LLE 算法中每个点的近邻权值在平移 , 旋转 , 伸缩变换下是保持不变的 .

• LLE 算法有解析的整体最优解 , 不需迭代 .

• LLE 算法归结为稀疏矩阵特征值计算 , 计算复杂度相对较小 , 容易执行 .

算法缺点• LLE 算法要求所学习的流形只能是不闭合的且在局部是线性的 .

• LLE 算法要求样本在流形上是稠密采样的 .

• LLE 算法中的参数 K, d 有过多的选择 .

• LLE 算法对样本中的噪音很敏感 .

Page 31: Manifold learning

05/03/2023 31

Laplacian Eigenmaps

• Using the notion of the Laplacian of a graph to compute a low-dimensional representation of the data– The laplacian of a graph is analogous to the Laplace Beltrami operator

on manifolds, of which the eigenfunctions have properties desirable for embedding (See M. Belkin and P. Niyogi for justification).

Page 32: Manifold learning

05/03/2023 32

Laplacian matrix (discrete Laplacian)

• Laplacian matrix is a matrix representation of a graph

– is the Laplacian matrix

– is the degree matrix

– is the adjacent matrix

Page 33: Manifold learning

05/03/2023 mlss09us_niyogi_belkin_gmml 33

Laplacian Eigenmaps

Page 34: Manifold learning

05/03/2023 mlss09us_niyogi_belkin_gmml 34

Laplacian Eigenmaps

Page 35: Manifold learning

05/03/2023 mlss09us_niyogi_belkin_gmml 35

Laplacian Eigenmaps

: Degree matrix

: (Weighted) adjacent matrix: Laplacian matrix

𝐷− 1𝐿𝑓=𝜆 𝑓

Page 36: Manifold learning

05/03/2023 36

Justification of optimal embedding

• We have constructed a weighted graph

• We want to map to a line so that connected points stay as close together as possible

• This can be done by minimizing the objective function

• It incurs a heavy penalty if neighboring points are mapped far apart.

Page 37: Manifold learning

05/03/2023 37

Justification of optimal embedding (Cont.)

Page 38: Manifold learning

05/03/2023 38

Justification of optimal embedding (Cont.)

The minimization problem reduces to finding

Note the constraint removes an arbitrary scaling factor in the embedding.

Using Lagrange multiplier and setting the derivative with respect to equal to zero, we obtain

The optimum is given by the minimum eigenvalue solution to the generalized eigenvalue problem (trivial solution: ).

Page 39: Manifold learning

05/03/2023 39

More methods for non-linear manifold learning

Page 40: Manifold learning

05/03/2023 40

Applications

• Super-resolution

• Laplacianfaces

Page 41: Manifold learning

05/03/2023 41

Super-Resolution Through Neighbor Embedding

• Intuition: small patches in the low- and high-resolution images form manifolds with similar local geometry in two distinct spaces.

• X: low-resolution image Y: target high-resolution image

• The algorithm is extremely analogous to LLE!– Step 1: construct neighborhood of each patch in X

– Step 2: compute the reconstructing weights of the neighbors that minimize the reconstruction error

– Step 3: perform high-dimensional embedding to (as opposed to the low-dimensional embedding of LLE)

– Step 4: Construct the target high-resolution image Y by enforcing local compatibility and smoothness constraints between adjacent patches obtained in step 3.

Chang, Hong, Dit-Yan Yeung, and Yimin Xiong. "Super-resolution through neighbor embedding." CVPR 2004.

Page 42: Manifold learning

05/03/2023 42

Super-Resolution Through Neighbor Embedding

• Training parameters– The number of nearest neighbors K

– The patch size

– The degree of overlap

Chang, Hong, Dit-Yan Yeung, and Yimin Xiong. "Super-resolution through neighbor embedding." CVPR 2004.

Page 43: Manifold learning

05/03/2023 43

Super-Resolution Through Neighbor Embedding

Chang, Hong, Dit-Yan Yeung, and Yimin Xiong. "Super-resolution through neighbor embedding." CVPR 2004.

Page 44: Manifold learning

05/03/2023 44

Laplacianfaces

• Mapping face images in the image space via Locality Preserving Projections (LPP) to low-dimensional face subspace (manifold), called Laplacianfaces.

• LLP is analogous to Laplacian Eigenmaps except the objective function– Laplacian Eigenmaps:

– LLP:

He, Xiaofei, et al. "Face recognition using laplacianfaces." Pattern Analysis and Machine Intelligence, IEEE Transactions on 27.3 (2005): 328-340.

Page 45: Manifold learning

05/03/2023 45

Laplacianfaces

Learning Laplacianfaces for Representation

1. PCA projection (kept 98 percent information in the sense of reconstruction error)

2. Constructing the nearest-neighbor graph

3. Choosing the weights

4. Optimize

The k lowest eigenvectors of

are choosing to form

is the so-called Laplacianfaces.

Page 46: Manifold learning

05/03/2023 46

Laplacianfaces

Two-dimensional linear embedding of face images by Laplacianfaces.

Page 47: Manifold learning

05/03/2023 47

Reference and Resources

• 浅谈流形学习 . Pluskid. 2010-05-29. http://blog.pluskid.org/?p=533&cpage=1

• Wikipedia: MDS, Manifold, Laplacian Matrix

• PCA: M. Bishop, PRML

• Eigenvalue decomposition and SVD: 机器学习中的数学(5)-强大的矩阵奇异值分解(SVD)及其应用• General Eigenvalue Problem: wolfram, tutorial

• Video lecture: Geometric Methods and Manifold Learning. Mikhail Belkin, Partha Niyogi (author of Laplacian eigenmap)

• MANI fold Learning Matlab Demo: http://www.math.ucla.edu/~wittman/mani/index.html

Page 48: Manifold learning

05/03/2023 48

Thank you.