Intrinsic stability-control method for recursive filters and neuralnetworks

摘要

Linear recursive filters can be adapted on-line but withninstability problems. Stability-control techniques exist, but they areneither computationally expensive or nonrobust. For the nonlinear ease,ne.g., locally recurrent neural networks, the stability ofninfinite-impulse response (IIR) synapses is often a condition to bensatisfied. This brief considers the knownnreparametrization-for-stability method for the on-line adaptation of IIRnadaptive filters. A new technique is also presented, based on thenfurther adaptation of the squashing function, which allows one tonimprove the convergence performance. The proposed method can be appliednto various filter realizations (direct forms, cascade or parallel secondnorder sections, lattice form), as well as to locally recurrent neuralnnetworks, such as the IIR multi-layer perceptron (IIR-MLP), withnimproved performance with respect to other techniques and to the case ofnno stability control. In this brief, the case of normalized latticenfilters is particularly considered; an analysis of the stabilizationneffects is also presented both analytically and experimentally