The Elastic Embedding Algorithm for Dimensionality Reduction Presenter

The Elastic Embedding Algorithm for Dimensionality Reduction Presenter : Wei-Hao Huang Authors : Miguel ´ A. Carreira-Perpi˜n´an ICML, 2010 Intelligent Database Systems Lab

Outlines n Motivation n Objectives n Methodology n Experiments n Conclusions n Comments Intelligent Database Systems Lab

Motivation • The disadvantage of dimensionality reduction – Difficult to understand their objective function. – Optimisation is costly and prone to local optima. Intelligent Database Systems Lab

Objectives • To propose a new dimensionality reduction ü More efficient and robust ü Further our understanding algorithms Intelligent Database Systems Lab

Methodology - Framework Objective function Laplacian eigenmaps High dimension dataset + Elastic Embedding SNE Low dimension data Intelligent Database Systems Lab

Methodology – Elastic Embedding • Object function • Gradient of E Intelligent Database Systems Lab

Methodology - Study of λ • N=2 • N>2 Intelligent Database Systems Lab

Methodology – Out of sample • Objective function • Mapping and reconstruction mappings Intelligent Database Systems Lab

Experiments – 2 D spiral Intelligent Database Systems Lab

Experiments – Swiss roll Intelligent Database Systems Lab

Experiments – COIL-20 dataset Intelligent Database Systems Lab

Conclusions • EE dimensionality reduction improves over SNE methods. • EE produces better quality more quickly and robustly. • All of ideas can be directly applied to SNE, t-SNE and earlier algorithms. Intelligent Database Systems Lab

Comments • Advantages – EE improves disadvantage of SNE on different versions • Applications – Dimensionality Reduction Intelligent Database Systems Lab