NEURAL CLASSIFIERS AND STATISTICAL PATTERN-RECOGNITION - APPLICATIONSFOR CURRENTLY ESTABLISHED LINKS

Recent research has linked back-propagation (BP) and radial-basis-func tion (RBF) network classifiers, trained by minimizing the standard mea n-square error (MSE), to two main topics in statistical pattern recogn ition (SPR), namely the Bayes decision theory and discriminant analysi s. However, so far, the establishment of these links has resulted in o nly a few practical applications for training, using, and evaluating t hese classifiers, This paper aims at providing more of these applicati ons, It first illustrates that while training a linear-output BP netwo rk, the explicit utilization of the network discriminant capability le ads to an improvement in its classification performance, Then, for lin ear-output BP and RBF networks, the paper defines a new generalization measure that provides information about the closeness of the network classification performance to the optimal performance, The estimation procedure of this measure is described and its use as an efficient cri terion for terminating the learning algorithm and choosing the network topology is explained, The paper finally proposes an upper bound on t he number of hidden units needed by an RBF network classifier to achie ve an arbitrary value of the minimized MSE, Experimental results are p resented to validate all proposed applications.