Blind Signal Processing by Complex Domain Adaptive Spline Neural Networks 2003英文电子书.pdf

IEEE TRANSACTIONS ON NEURAL NETWORKS, VOL. 14, NO. 2, MARCH 2003 399 Blind Signal Processing by Complex Domain Adaptive Spline Neural Networks Aurelio Uncini, Member, IEEE, and Francesco Piazza, Member, IEEE Abstract—In this paper, neural networks based on an adap- In both real and complex cases, however, the cumulative den- tive nonlinear function suitable for both blind complex time do- sity functions of the sources are usually unknown. Although main signal separation and blind frequency domain signal decon- simulations often exhibit good results also for nonlinearities that volution, are presented. This activation function, whose shape is modified during learning, is based on a couple of spline functions, do not exactly match the signals, in the general case a better esti- one for the real and one for the imaginary part of the input. The mation of the true nonlinear functions can be important. There- shape control points are adaptively changed using gradient-based fore, several approaches have been proposed in the real case to techniques. B-splines are used, because they allow to impose only obtain flexible nonlinearities (see for example [10]–[14]), which simple constraints on the control parameters in order to ensure a allow to optimize the shape of the functions with respect to the monotonously increasing characteristic. This new adaptive func- tion is then applied to the outputs of a one-layer neural network input signals. in order to separate complex signals from mixtures by maximizing In [10] and [11], an interesting technique that uses mixture the entropy of the function outputs. We derive a simple form of the models for the density fun