A novel approach to the convergence of neural networks for signal processing

Abstract

Summary form only given. A novel deterministic approach to the convergence of (stochastic) learning algorithms is presented. The link is the new concept of time-average invariance which is a property of deterministic signals but resembles the realizations of stochastic signals that are ergodic and stationary. An unsupervised learning algorithm is considered. Signals are viewed as deterministic functions, but satisfy a property called time-average invariance. As such, deterministic-based analysis can be applied to stochastic-like signals. Consequently, the complexity of the convergence analysis is significantly reduced.<>