Non-linear Dimensionality Reduction by Locally Linear Inlaying Presented by Peng Zhang Tianjin University, China Yuexian Hou, Peng Zhang, Xiaowei Zhang, Xingxing Xu, & Wenjie Li
Why Nonlinear Methods Inherent dimensionalities != the linear transformation of initial features Face image Example: Figures from ISOMAP paper L
Manifold Learning Local Linearity Global Embedding
A Run of ISOMAP L PCA ISOMAP Original
Another Run of ISOMAP Figures from ISOMAP paper
Other classical NLDR methods LLE LTSA LLC Laplacian Eigenmaps LPP ……
Drawbacks of these methods Most of these methods at least quadratic time in N fails on non-uniform dataset. not robust to heterogeneous noise
Locally Linear Inlaying (LLI) Linear time cost Perform well on nonconvex samples robust to heterogeneous noise
1. Running time comparison
2. Non-convex Swiss Roll Dataset L
3. Swiss roll with heterogeneous noise Isomap LTSA LLC LLI Original 2-D projection
4. “ Frey Face ” dataset Intensity of Illumination Pose
Connections to IR LLI in text background? Positive Linearity in local area is better than global data Obtain local document sets (clusters or subtopics) Reasons for Negative Different metric in text and imageDifferent metric Image have natural manifold meaningmanifold meaning LLI in Renaissance Project? Local context basis Global context-sensitive embedding
Thank You! Any Question???