A MODEL-BASED NEURAL-NETWORK FOR TRANSIENT SIGNAL-PROCESSING

A hierarchical maximum Likelihood Adaptive Neural System (MLANS) is pr oposed for transient signal processing. This is a new type of neural n etwork that incorporates a model-based concept, leading to greatly inc reased learning efficiency compared to conventional, nonparametric neu ral networks. This neural network has a hierarchical, two-layer struct ure: the bottom, signal modeling layer and the top, classification lay er. The signal modeling layer of MLANS, which is a subject of the curr ent publication, operates on a two-dimensional representation of the s ignal such as the short-term spectrum or the Wigner transform domain. The internal model of a signal is developed by MLANS from a single exa mple. From an estimation theory point of view, MLANS performs a novel type of spectrum estimation, which is significantly more efficient tha n the classical minimum entropy or maximum likelihood spectrum estimat ors for AR signals. In addition, the MLANS model parsimoniously uses p arameters in a time-frequency domain, reducing a number of parameters required for an accurate signal characterization hy several orders of magnitude, compared to existing speech signal characterization techniq ues. The MLANS classification layer performs the maximum likelihood es timation of statistical mixture model parameters. The learning efficie ncies of each layer approach the information-theoretic limits determin ed by the Cramer-Rao bound.