Generalized SOM with Application to Facial Gender Identification
This paper generalizes the self-organizing map (SOM) in flat Euclidean space with the aid of Riemannian exponential and logarithmic maps and obtains the generalized self-organizing map (G-SOM) on Riemannian manifold. Both sequential and batch learning algorithms for the G-SOM are presented. G-SOM can preserve the intrinsic topological neighborhood structure of the input patterns on output space. Theoretical analysis shows that G-SOM will outperform SOM when the input space is a non-linear manifold. Gender identification experiments on FERET support our theoretical results.
NETL
NSTL
万方数据
维普
