Department of
Epidemiology &

Genetic Epidemiology

Global Health

Behavior and Prevention

Health Care

Modern Biostatistics

Biostatistics Seminar

Thursday, October, 17, 2013
12:00pm-1:00pm -- WG73
Bring your own lunch
Water and light refreshments (compliments of some faculty) will be served

Title:rNMF: Modern Dimension Reduction and Robust Image Compression and Applications

Speaker: Dr. Yifan Ethan Xu, Department of Epidemiology and Biostatistics, Case Western Reserve University


Data sets with large number of observations and/or features appear frequently in many areas. Non-negative matrix factorization (NMF) is a modern method for obtaining low dimensional structure from high dimensional data, and it has been successfully applied in a wide range of fields including computational biology (e.g. DNA microarrays analysis, molecular pattern discovery, etc.), image processing, and document clustering. Comparing to traditional methods such as principle component analysis (PCA), NMF decompose non-negative data matrix into non-negative linear combinations of basis, which reveals parts-based underling structure that is intuitive to interpret. As with other dimension reduction methods, NMF is sensitive to outliers and large noises. To solve the problem we developed a comprehensive robust NMF package (rNMF) that compresses noisy data while simultaneously control the sparsity of the decomposition. It allows both automatic and semi-supervised control to handle different types of corruptions.

In this talk, I will introduce regular and robust NMF algorithms. The advantages and limitations of the procedures will be discussed. The methods are illustrated with examples in biological data decomposition and simulated corrupted tumor image compression. A multilevel image compression and cleaning procedure that combines rNMF and human control will also be shown. The talk is based on a joint work with Jiayang Sun (CWRU), Kenneth Lopiano (SAMSI) and Stanley Young (NISS).

Hosting: Ilimitada Hosting Chile VPS