Post on 12-Aug-2015
Non-negative Matrix Factorization with
Sparseness Constraints
Patrik O. Hoyer
HIIT Basic Research Unit
Department of Computer Science
University of HelsinkiFinland
Topics
• Introduction
• Problem overview
• Overall methodology
• Integrating NMF with Age estimation problem
• Extended NMF
• Result found for age estimation
• Conclusion
Non-negative Matrix Factorization
• Factor A = WH• A – matrix of data• m non-negative scalar variables • n measurements form the columns of A
• W – m x r matrix of “basis vectors”• H – r x n coefficient matrix• Describes how strongly each building block is
present in measurement vectors
Non-negative Matrix Factorization
Non-negative Matrix Factorization
• The image database is regarded as an n*m matrix V,each column of which contains n non-negative pixel values of one of the m facial images
• Our purpose is to construct approximate factorizations of the form,
where each column of H contains the coefficient vector ht corresponding to the measurement vector, Vt
• Written in this form, it becomes apparent that a linear data representation is simply a factorizationof the data matrix
• Given a data matrix V, the optimal choice of matrices W and H are to be defined to be those nonnegative matrices that minimize the reconstruction error between V and WH
• Various error functions have been proposed (Paatero and Tapper, 1994; Lee and Seung, 2001), perhaps the most widely
used one is eucledian distance measure
•
.
Adding sparseness constrains to NMF
• The concept of ‘sparse coding’ refers to a representational scheme where only a few units (out of alarge population) are effectively used to represent typical data vectors
• This implies most units taking values close to zero while only few take significantly non-zero values.
• On a normalized scale, the sparsest possible vector (only a single component is non-zero) should have a sparseness of one, whereas a vector with all elements equal zero should have a sparseness of zero.
• In this paper, we use a sparseness measure based on the relationship between the L1 norm andthe L2 norm:
NMF with sparseness component
• Our aim is to constrain NMF to find solutions with desired degrees of sparseness.
• What exactly should be sparse? The basis vectors W or the coefficients H? Depends on scenario
• When trying to learn useful features from a database of images, it might make sense to require bothW and H to be sparse, pointing that any given object is present in few images and affects only asmall part of the image.
Age Estimation Based on Extended Non-negative
Matrix Factorization
Ce Zhan, Wanqing Li, and Philip OgunbonaSchool of Computer Science and Software Engineering
University of Wollongong, Australia
Proposed method : Extended NMF
• Extended NMF (ENMF) impose orthogonality constraint on basis matrix W while controlling the sparseness of coefficient matrix H.
• To reduce the overlapping between basis images, different bases should be asorthogonal as possible so as to minimize the redundancy
• Denote : U=WTW
• The orthogonality constraint can be imposed by minimizing
• Maximum sparsity in the coefficient matrix makes sure that a basis componentcannot be further decomposed into more components, thus the overlapping between basis images is further re-duced.
• We want :
Extended NMF
• The objective function of the proposed method would be:
• β is a small positive constant
ENMF is defined as following optimization problem
Experimental result
Experimental result