A Quadratic Divergence-Based Independence Measure Applied to Linear-Quadratic Mixtures
Denis G. Fantinato, Rafael A. Ando, Aline Neves, Leonardo T. Duarte, Christian Jutten, Romis Attux

DOI: 10.14209/sbrt.2016.50
Evento: XXXIV Simpósio Brasileiro de Telecomunicações (SBrT2016)
Keywords: Blind Source Separation Linear Quadratic Quadratic Divergence Kernel Methods
Abstract
In the context of the Blind Source Separation (BSS) problem, the use of Mutual Information (MI) as an independence measure can be very effective, even for certain types of nonlinear mixtures. However, in this case, it is generally necessary to estimate the joint and the marginal distributions associated with the random variables of interest. In this work, we consider the kernel methods for distribution estimation allied to a convenient metric as alternative to classical MI: the quadratic divergence. The proposed method is applied to the problem of Linear-Quadratic mixtures, using a recurrent network for separation and an evolutionary algorithm, the opt-aiNet, for parameter optimization. Its performance is analyzed and compared with the classical MI estimated via histogram in three different scenarios including synthetic and real data. The results are favorable to the proposal, especially for small data sets.

Download