Nonlinear Dimensionality Reduction. Jose A. Costa

Similar documents
Non-linear Dimensionality Reduction

Nonlinear Methods. Data often lies on or near a nonlinear low-dimensional curve aka manifold.

Laplacian Eigenmaps for Dimensionality Reduction and Data Representation

Data-dependent representations: Laplacian Eigenmaps

Intrinsic Structure Study on Whale Vocalizations

Robust Laplacian Eigenmaps Using Global Information

Unsupervised dimensionality reduction

Manifold Learning and it s application

Unsupervised Learning Techniques Class 07, 1 March 2006 Andrea Caponnetto

Nonlinear Dimensionality Reduction

CSE 291. Assignment Spectral clustering versus k-means. Out: Wed May 23 Due: Wed Jun 13

Laplacian Eigenmaps for Dimensionality Reduction and Data Representation

Graph-Laplacian PCA: Closed-form Solution and Robustness

A Duality View of Spectral Methods for Dimensionality Reduction

Learning Eigenfunctions: Links with Spectral Clustering and Kernel PCA

A Duality View of Spectral Methods for Dimensionality Reduction

Distance Metric Learning in Data Mining (Part II) Fei Wang and Jimeng Sun IBM TJ Watson Research Center

Diffusion Wavelets and Applications

Manifold Learning: Theory and Applications to HRI

Learning a Kernel Matrix for Nonlinear Dimensionality Reduction

Graphs, Geometry and Semi-supervised Learning

Connection of Local Linear Embedding, ISOMAP, and Kernel Principal Component Analysis

Data dependent operators for the spatial-spectral fusion problem

EECS 275 Matrix Computation

Discriminant Uncorrelated Neighborhood Preserving Projections

Learning a kernel matrix for nonlinear dimensionality reduction

Lecture 10: Dimension Reduction Techniques

Face Recognition Using Laplacianfaces He et al. (IEEE Trans PAMI, 2005) presented by Hassan A. Kingravi

Dimensionality Reduc1on

Apprentissage non supervisée

Nonlinear Dimensionality Reduction

Dimensionality Reduction AShortTutorial

Informative Laplacian Projection

Global (ISOMAP) versus Local (LLE) Methods in Nonlinear Dimensionality Reduction

Riemannian Manifold Learning for Nonlinear Dimensionality Reduction

Nonlinear Dimensionality Reduction by Semidefinite Programming and Kernel Matrix Factorization

Locally Linear Embedded Eigenspace Analysis

Dimensionality Reduction:

Localized Sliced Inverse Regression

Dimension Reduction and Low-dimensional Embedding

Large-Scale Manifold Learning

Machine Learning. Data visualization and dimensionality reduction. Eric Xing. Lecture 7, August 13, Eric Xing Eric CMU,

Focus was on solving matrix inversion problems Now we look at other properties of matrices Useful when A represents a transformations.

Principal Component Analysis (PCA)

TANGENT SPACE INTRINSIC MANIFOLD REGULARIZATION FOR DATA REPRESENTATION. Shiliang Sun

Bi-stochastic kernels via asymmetric affinity functions

Dimensionality Reduction: A Comparative Review

Clustering in kernel embedding spaces and organization of documents

Locality Preserving Projections

Spectral Methods for Semi-supervised Manifold Learning

LECTURE NOTE #11 PROF. ALAN YUILLE

Is Manifold Learning for Toy Data only?

Learning on Graphs and Manifolds. CMPSCI 689 Sridhar Mahadevan U.Mass Amherst

Dimension reduction methods: Algorithms and Applications Yousef Saad Department of Computer Science and Engineering University of Minnesota

Machine Learning. CUNY Graduate Center, Spring Lectures 11-12: Unsupervised Learning 1. Professor Liang Huang.

Statistical Pattern Recognition

Iterative Laplacian Score for Feature Selection

Advances in Manifold Learning Presented by: Naku Nak l Verm r a June 10, 2008

MANIFOLD LEARNING: A MACHINE LEARNING PERSPECTIVE. Sam Roweis. University of Toronto Department of Computer Science. [Google: Sam Toronto ]

Nonlinear Manifold Learning Summary

Manifold Regularization

Statistical and Computational Analysis of Locality Preserving Projection

Dimensionality Reduction: A Comparative Review

Theoretical analysis of LLE based on its weighting step

Mid-year Report Linear and Non-linear Dimentionality. Reduction. applied to gene expression data of cancer tissue samples

Regression on Manifolds Using Kernel Dimension Reduction

Improved Local Coordinate Coding using Local Tangents

Dimension Reduction Techniques. Presented by Jie (Jerry) Yu

Graph Metrics and Dimension Reduction

Diffusion Geometries, Global and Multiscale

Lecture 3: Compressive Classification

Advanced Machine Learning & Perception

Nonlinear Learning using Local Coordinate Coding

L26: Advanced dimensionality reduction

Learning gradients: prescriptive models

Supplemental Materials for. Local Multidimensional Scaling for. Nonlinear Dimension Reduction, Graph Drawing. and Proximity Analysis

Gaussian Process Latent Random Field

Fisher s Linear Discriminant Analysis

A Tour of Unsupervised Learning Part I Graphical models and dimension reduction

Lecture: Some Practical Considerations (3 of 4)

Manifold Regularization

LEC 2: Principal Component Analysis (PCA) A First Dimensionality Reduction Approach

An Empirical Comparison of Dimensionality Reduction Methods for Classifying Gene and Protein Expression Datasets

Dimensionality Estimation, Manifold Learning and Function Approximation using Tensor Voting

Statistical Learning. Dong Liu. Dept. EEIS, USTC

Image Analysis & Retrieval Lec 13 - Feature Dimension Reduction

ISSN: (Online) Volume 3, Issue 5, May 2015 International Journal of Advance Research in Computer Science and Management Studies

Spectral Dimensionality Reduction

A graph based approach to semi-supervised learning

Classification of handwritten digits using supervised locally linear embedding algorithm and support vector machine

Statistical Machine Learning

Semi Supervised Distance Metric Learning

PARAMETERIZATION OF NON-LINEAR MANIFOLDS

Certifying the Global Optimality of Graph Cuts via Semidefinite Programming: A Theoretic Guarantee for Spectral Clustering

MACHINE LEARNING. Methods for feature extraction and reduction of dimensionality: Probabilistic PCA and kernel PCA

Beyond the Point Cloud: From Transductive to Semi-Supervised Learning

Carlo Vittorio Cannistraci. Minimum Curvilinear Embedding unveils nonlinear patterns in 16S metagenomic data

Relevance Aggregation Projections for Image Retrieval

SINGLE-TASK AND MULTITASK SPARSE GAUSSIAN PROCESSES

ECE 521. Lecture 11 (not on midterm material) 13 February K-means clustering, Dimensionality reduction

Transcription:

Nonlinear Dimensionality Reduction Jose A. Costa Mathematics of Information Seminar, Dec.

Motivation Many useful of signals such as: Image databases; Gene expression microarrays; Internet traffic time series; etc. are very high dimensional in nature. Curse of dimensionality hinders the analysis of such datasets: 1. Poor statistical performance; 2. Unmanageable computational complexity.

Motivation However, the apparent high complexity of such signals is often just an artifact of the measurement process and its data representation. not all inputs carry independent information! 1 1 1 1 1 2 2 2 2 2 1 2 1 2 1 2 1 2 1 2 1 1 1 1 1 2 2 2 2 2 1 2 1 2 1 2 1 2 1 2 28x28 pixels images embedded in

Motivation Approach: design dimensionality reduction algorithms that learn compact representations of high dimensional data. For what purpose? 1 1 1 1 2 2 2 2 1 2 1 2 1 2 1 2 1 2 3 Estimate parameters that produced the data set Classify data

Outline 1. Finding Linear Data Representations: i. Principal Component Analysis (PCA) ii. Multidimensional Scaling (MDS) 2. From Linear to Nonlinear: MDS to ISOMAP 3. Spectral Graph Methods: Laplacian Eigenmaps 4. Adding Additional Constraints: The Supervised and Semi-Supervised Learning Problems.

Setup Assumption: The set of data points lives on a compact m-dimensional manifold 1 Swiss roll 2D manifold in

Manifold Learning Manifold learning problem setup: Input: a finite sampling of a m-dimensional manifold. Output: embedding of into a subset (usually ) without any prior knowledge about.

Background on Manifold Learning Reconstructing the mapping and attributes of the manifold from a finite dataset falls into the general manifold learning problem. Manifold reconstruction: 1. ISOMAP, Tenenbaum, de Silva, Langford (); 2. Locally Linear Embeddings (LLE), Roweiss, Saul (); 3. Laplacian Eigenmaps, Belkin, Niyogi (2); 4. Hessian Eigenmaps (HLLE), Grimes, Donoho (3);. Local Space Tangent Alignment (LTSA), Zhang, Za (3); 6. SemiDefinite Embedding (SDE), Weinberger, Saul (4), Sun, Boyd, Xiao, Diaconis (4).

Principal Component Analysis Given n data points,, what is the best approximating linear subspace of dimension m?

Principal Component Analysis 2. PCA as an optimization problem: a. objective function: where is the Graph Laplacian.

Manifold Learning and Classification 1 8 points uniform on Swiss roll, 4 each class

Manifold Learning and Classification Popular manifold learning algorithms: 1 1.6.4.2.2.4.6 1 6 4 4 6 ISOMAP.8.8.6.4.2.2.4.6 Laplacian Eigenmaps

Laplacian Eigenmaps Laplacian Eigenmaps: preserving local information (Belkin & Niyogi 2) 1. Constructing an Adjacency Graph: a. compute a k-nn graph on the dataset; b. compute a similarity/weight matrix W between data points, that encodes neighborhood information (e.g., heat kernel):

Laplacian Eigenmaps 2. Manifold learning as an optimization problem: a. objective function: where is the Graph Laplacian. b. embedding is solution of ( )

Laplacian Eigenmaps 3. Eigenmaps: a. solution to ( ) is given by the m generalized eigenvectors associated with the m smallest generalized eigenvalues that solve : or equivalent eigenvectors of the normalized Graph Laplacian b. if is the collection of such eigenvectors, then the embedded points are given by

Manifold Learning and Classification Adding class dependent constraints virtual class vertices. 1

Manifold Learning and Classification 1. If C is the class membership matrix (i.e., c_ij = 1 if point j is from class i), define the objective function: where centers and, are the virtual class is a regularization parameter. 2. Embedding is now solution of where L is Laplacian of augmented weight matrix

Manifold Learning and Classification 1.6.4.2.2.4.6 1 1 6 4 4 6.8.8.6.4.2.2.4.6 ISOMAP Laplacian Eigenmaps.3.2.2.1.1.. Classification Constrained Dimensionality Reduction.1.1.2.2.1.1...1.1.2

Manifold Learning and Classification Error rates for k-nn classifier using pre-processing dimensionality reduction versus full dimensional data

Manifold Learning and Classification Semi-Supervised Learning on Manifolds unlabeled samples labeled samples 1

Manifold Learning and Classification Algorithm: 1. Compute the constrained embedding of the entire data set, inserting a zero column in C for each unlabeled sample. labeled samples unlabeled samples

Manifold Learning and Classification 2. Fit a (e.g., linear) classifier to the labeled embedded points by minimizing a loss function (e.g., quadratic): 3. For an unlabeled point, label it using the fitted (linear) classifier:

Manifold Learning and Classification 6 Error rate (%) 4 3 k NN Laplacian CCDR 3 4 6 7 Number of labeled points Percentage of errors for labeling unlabeled samples as a function of the number of labeled points, out of a total of points on the Swiss roll.

Summary and Ongoing Work 1. Preservation of local geometric structures and class label information; 2. Optimization problem with global minimum; 3. Applicable to both supervised and semi-supervised learning paradigms; Out-of-sample extension; Connections between classification, dimensionality reduction and dimensionality expansion via kernel methods