Comparative Study of the Practical Characteristics of Neural Network andConventional Pattern Classifiers

摘要

Six pattern classifiers were implemented on a serial computer and compared usingartificial and speech recognition tasks. Two neural network classifiers (radial basis function and high-order polynomial GMDH network) and four more conventional classifiers (Gaussian mixture, linear tree, KD-tree, and condensed K nearest neighbor) were investigated. These classifiers were chosen to be representative of different approaches to the task of pattern recognition, and to complement and extend those that were investigated in a previous study. The goal was to analyze and compare the performance and behavior of the classifiers on different tasks in terms of classification error rate, complexity, memory requirements, and training and classification times. Radial basis function classifiers generalized well to high-dimensional spaces, and provided low error rates with training times that were much less than those of back-propagation classifiers. High-order polynomial classifiers provided intermediate error rates, but often required very long training times. In addition, the decision regions formed did not extrapolate well to data-poor regions of the input space.