Nonlinear dynamics of feedback multilayer perceptrons.

We study the nonlinear dynamics of multilayer perceptrons with feedback and propose their application to the analysis of signals with complex time dependence. We show that their dynamics provides a built-in time-warping invariance, as, e.g., required for presentation speed fluctuations in speech recognition. We suggest an appropriate learning rule (open-loop learning), give an analytical stability condition for the resulting multistable states, and determine their basins of attraction. To demonstrate their utility for possible applications, we consider the example of a three-stage feedback multilayer perceptron that is trained to detect words in a sequence of letters and does it with perfect invariance with respect to presentation speed fluctuations.