Improved singular value decomposition by using neural networks

In this paper dedicated parallel processing techniques based on neural networks for iterative computation of singular value decomposition (SVD) in a real-time signal processing environment is explored. Recently proposed linear neuron multilayer feedforward networks for SVD are investigated in relation to well-known existing numerical SVD techniques. A new modified feedforward neural architecture with fewer feedforward neuron layers is presented which is found to be more suitable for the SVD of ill-conditioned matrices compared to Cichocki´s networks. Also, we have introduced an appropriate method for evaluating different neural SVD architectures based on their numerical stability