This project will attempt to develop more advanced designs for artifical neural networks which can learn from experience how to predict of filter streams of inputs over time. The work will build upon novel designs developed under a previous NSF grant, which are being tested in applications such as blind deconvolution and interference cancelling in nonstationary communication channels (e.g., making cellular phones work better), identification and control of nonlinear plants, prediction of financial time-series and classification of time varying patterns such as speech recognition and transient signal processing (radar, sonar, biological). The basic approach will be to deepen the theoretical understanding of the limitations of the existing designs in these applications, and to develop new general-purpose designs--rooted in an exploration of basic theory--to overcome these limitations. Among the theoretical issues given priority are the problem of mixing discrete and continuous variables in the context, the problem of time-warping (e.g. different speakers speaking at different speeds), and the problems of robustness over time.
