Abstract
In analog implementations of adaptive algorithms, limited voltage ranges and inevitable dc, offset voltages that arise in the learning circuit seriously degrade the convergence performance. Using the circuit structures proposed in this paper, these problems are eliminated and the learning characteristic of the resulting network is improved. The test results are obtained from a structure with two channels using a stochastic gradient algorithm in order to minimize the cost function defined as the output signal cross-correlation. This algorithm allows the identification of the channels and the reconstruction of the signals so long as the signals are L2 integrable.