A New Connectionist Model for Time-Varying Signal Classification

点击翻译按钮获取中文摘要

This project will seek a unifying connectionist framework for the classification of time varying signals based on the gamma model proposed by (de Vries and Principe, 1990, 1991, 1992). The gamma model extends the conventional additive model because the multiplication of output activations with weights is subsituted by a temporal convolution, yielding a computationally improved additive model. This research will have theoretical and practical components. New insights on the role of temporal weights will be sought, using the theory of recursive estimation. Likewise, a paradigm to study how the neural network is representing the history of the input signal will be sought, using Hilbert spaces. The gamma memory projects a vector of unconstrained dimension onto a predefined subspace. This same principle will be extended to learning, producing a backpropagation through time algorithm that has memory requirements independent of the length of the input signal. As an application, the gamma network will be implemented in a multiprocessor system for speech recognition. The P.I. will contrast the gamma model with two of the best available paradigms, Lang's time delay neural network and Tank's and Hopfield concentration in time networks, in terms of performance, network complexity and learning effectiveness.