Analogue hardware and some convergence properties of the sources separation algorithm

We present in our paper a simplified analogue hardware of the two-neuron Herault-Jutten network (1989, 1991) for separation of two sources from a linear and instantaneous mixture. The simplification is in the choice of the nonlinear functions used by the learning rule. Based on theoretical considerations and simulation results, the influence of these nonlinearities and the statistical nature of sources on the convergence of the algorithm are pointed out. In order to determine the properties and the limitations of our analogue hardware, the behavior of the algorithm is derived finally for strong nonlinear functions, as they are actually implemented in our circuit.