26
Manifold learning Jan Kamenický

Manifold learning

  • Upload
    quinta

  • View
    65

  • Download
    1

Embed Size (px)

DESCRIPTION

Manifold learning. Jan Kamenický. Nonlinear dimensionality reduction. Many features ⇒ many dimensions Dimensionality reduction Feature extraction (useful representation) Classification Visualization. Manifold learning. WhaT maniFold ? - PowerPoint PPT Presentation

Citation preview

Page 1: Manifold learning

Manifold learningJan Kamenický

Page 2: Manifold learning

Many features ⇒ many dimensions

Dimensionality reduction◦ Feature extraction (useful representation)◦ Classification◦ Visualization

Nonlinear dimensionality reduction

Page 3: Manifold learning

WhaT maniFold?◦ Low dimensional embedding of high dimensional

data lying on a smooth nonlinear manifold

Linear methods fail◦ i.e. PCA

Manifold learning

Page 4: Manifold learning

Unsupervised methods◦ Without any a priori knowledge

ISOMAPs◦ Isometric mapping

LLE◦ Locally linear embedding

Manifold learning

Page 5: Manifold learning

Core idea◦ Use geodesic distances on the manifold instead of

Euclidean

Classical MDS◦ Maps data to the lower dimensional space

ISOMAP

Page 6: Manifold learning

Select neighbours◦ K-nearest neighbours◦ ε-distance neighbourhood

Create weighted neighbourhood graph◦ Weights = Euclidean distances

Estimate the geodesic distances as shortest paths in the weighted graph◦ Dijkstra’s algorithm

Estimating geodesic distances

Page 7: Manifold learning

Dijkstra’s algorithm 1) Set distances (0 for initial, ∞ for all other nodes),

set all nodes as unvisited 2) Select unvisited node with smallest distance as

active 3) Update all unvisited neighbours of the active

node (if the computed distance is smaller) 4) Mark active node as visited (it has now minimal

distance), repeat from 2) as necessary

Page 8: Manifold learning

Time complexity◦ O(|E|dec+|V|min)

Implementation◦ Sparse edges◦ Fibonacci heap as a priority queue◦ O(|E|+|V|log|V|)

Geodesic distances in ISOMAP◦ O(N2logN)

Dijkstra’s algorithm

Page 9: Manifold learning

Input◦ Dissimilarities (distances)

Output◦ Data in a low-dimensional embedding, with

distances corresponding to the dissimilarities

Many types of MDS◦ Classical◦ Metric / non-metric (number of dissimilarity

matrices, symmetry, etc.)

Multidimensional scaling (MDS)

Page 10: Manifold learning

Quantitative similarity Euclidean distances (output) One distance matrix (symmetric)

Minimizing the stress function

Classical MDS

Page 11: Manifold learning

We can optimize directly◦ Compute double-centered distance matrix

◦ Note:

◦ Perform SVD of B

◦ Compute final data

Classical MDS

Page 12: Manifold learning

Covariance matrix

Projection of centered X onto eigenvectors of NS (result of the PCA of X)

MDS and PCA correspondence

Page 13: Manifold learning

ISOMAP

Page 14: Manifold learning

ISOMAP

Page 15: Manifold learning

How many dimensions to use?◦ Residual variance

Short-circuiting◦ Too large neigbourhood (not enough data)◦ Non-isometric mapping◦ Totally destroys the final embedding

ISOMAP

Page 16: Manifold learning

Conformal ISOMAP◦ Modified weights in geodesic distance estimate:

◦ Magnifies regions with high density◦ Shrinks regions with low density

ISOMAP modifications

Page 17: Manifold learning

C-ISOMAP

Page 18: Manifold learning

Landmark ISOMAP◦ Use only geodesic distances from several

landmark points (on the manifold)◦ Use Landmark-MDS for finding the embedding

Involves triangulation of non-landmark data◦ Significantly faster, but higher chance for “short-

circuiting”, number of landmarks has to be chosen carefully

ISOMAP modifications

Page 19: Manifold learning

Kernel ISOMAP◦ Ensures that the B (double-centered distance

matrix) is positive semidefinite by constant-shifting method

ISOMAP modifications

Page 20: Manifold learning

Core idea◦ Estimate each point

as a linear combination of it’s neighbours – find best such weights

◦ Same linear representation will hold in the low dimensional space

Locally linear embedding

Page 21: Manifold learning

Find weights Wij by constrained minimization

Neighbourhood preserving mapping

LLE

Page 22: Manifold learning

Low dimensional representation Y

We take eigenvectors of M corresponding to its q+1 smallest eigenvalues

Actually, different algebra is used to improve numeric stability and speed

LLE

Page 23: Manifold learning

LLE

Page 24: Manifold learning

LLE

Page 25: Manifold learning

ISOMAP◦ Preserves global geometric properties (geodesic

distances), especially for faraway points

LLE◦ Preserves local neighbourhood correspondence

only◦ Overcomes non-isometric mapping◦ Manifold is not explicitly required◦ Difficult to estimate q (number of dimensions)

ISOMAP vs LLE

Page 26: Manifold learning

The end