Title :
Unconstrained functional criteria for canonical correlation analysis
Author :
Hasan, Mohammed A.
Author_Institution :
Dept. of Electr. & Comput. Eng., Minnesota Univ., Duluth, MN, USA
Abstract :
Unconstrained and constrained optimization criteria for extracting multiple true canonical variates and canonical correlations are proposed. These include a weighted information criterion (WINC-CCA) and a shifted information criterion (SINC-CCA) for searching the optimal solutions. The gradient flows of the proposed CCA functions are analyzed and some of their convergence properties are presented. The main feature of this approach is that the actual canonical variates are automatically obtained. Based on the gradient-ascent method, many algorithms for performing the true CCA recursively are provided.
Keywords :
convergence; correlation methods; gradient methods; information theory; optimisation; recursive estimation; SINC-CCA; WINC-CCA; canonical correlations; constrained optimization criteria; convergence; gradient-ascent method; multiple canonical variate parallel extraction; multiple true canonical variates; recursive CCA; shifted information criterion; unconstrained optimization criteria; weighted information criterion; Adaptive algorithm; Concurrent computing; Constraint optimization; Data mining; Kernel; Matrix decomposition; Singular value decomposition; Statistics; Symmetric matrices; Vectors; Fast adaptive algorithms; canonical correlation extraction; information criteria;
Conference_Titel :
Circuits and Systems, 2005. ISCAS 2005. IEEE International Symposium on
Print_ISBN :
0-7803-8834-8
DOI :
10.1109/ISCAS.2005.1465088
