An auto-adaptive synthetic neural network for real-time separation of independent signal sources

Discusses the following classical signal processing problem: given N physically distinct measurements which represent a priori unknown linear combinations of N independent signal sources, the network autoadaptively extracts the original independent signals. The authors consider the N input/output case. The analysis is restricted to signals which are zero mean, stationary and either periodic or aperiodic, deterministic or nondeterministic and to a medium which is homogeneous and introduces no time delay. In practice, the source separation algorithm also works well for nonstationary signals. The authors investigated the constraints which the network's learning rule must satisfy and tested the proposed learning rules with digital simulations. To achieve real-time operation, they implemented three different circuit designs for 2 and 6 input/output networks using 2 mu m n-well analog VLSI hardware.<<ETX>>