Nonlinearly Activated Neural Network For Solving Time-varying Complex Pseudo-inverses

In this paper,we investigate a new complex-valued ZNN,which is called complex-valued ZNN,for pseudo-inverse of time-varying matrices.The complex-valued ZNN is proposed based on a matrix-valued error function in the domain of complex numbers.Utilizing the complexvalued first-order time-derivative,the complex-valued ZNN model with the proposed nonlinear complex-valued activation functions,can globally exponentially convergence to the theoretical inverse in an error-free manner.We introduce the building of neural network to calculate the time-varying matrix pseudo-inverse.The neural network calculate time-varying matrix pseudo-inverse from time-varying matrix inverse,which analyze real-valued and complexvalued.The activation function of neural network includes two kinds of function which called H function.One of the function is used to calculate real and imaginary parts of entry value,another is used to calculate modulus and arguments of entry value.In the same time,the convergence of activation function is certified.At the end of article,we use the Matlab to analog simulate the neural network,which calculate time-varying matrix pseudo-inverse.The activation function is nonlinear activation function,which is power function,power-sigmoid function,bipolar-sigmoid function,bipolar-power function in the time-varying matrix inverse.And we verify the neural network is bounded under the activation function which is selected bipolar-power function.The activation function is power function,power-sigmoid function,bipolarsigmoid function in the calculation of time-varying matrix pseudo-inverse.We also verify the power-sigmoid function has better convergence than other function in the neural network.