Nonlinear independent component analysis based on interval optimization

In this paper, a nonlinear independent component analysis method is developed. It is at heart a smooth mapping method, also incorporates interval optimization to improve the learning performance. The nonlinear inverse mapping from observations to estimated source signals is modeled by a multilayer perceptron network, while their weights are replaced by interval numbers. Then a Branch-and-Bound optimization algorithm is constructed by using interval analysis. It can overcome the problems of easily get trapped in the local minimum and unsatisfactory convergence speed, which would otherwise be severed in unsupervised learning with nonlinear models. The experimental results show that the proposed algorithm can efficiently separate both the same type and different type sources only from observations.