Nearest matched filter classification of spatiotemporal patterns

Abstract

Recent advances in massively parallel optical and electronic neural network processing technology have made it plausible to consider the use of matched filter banks containing large numbers of individual filters as pattern classifiers for complex spatiotemporal pattern environments such as speech, sonar, radar, and advanced communications. This paper begins with an overview of how neural networks can be used to approximately implement such multidimensional matched filter banks. The nearest matched filter classifier is then formally defined. This definition is then reformulated to show that the classifier is equivalent to a nearest neighbor classifier in a separable infinite-dimensional metric space that specifies the local-in-time behavior of spatiotemporal patterns. The result of Cover and Hart is then applied to show that, given a statistically comprehensive set of filter templates, the nearest matched filter classifier will have near-Bayesian performance for spatiotemporal patterns. The combination of near-Bayesian classifier performance with the excellent performance of matched filtering in noise yields a powerful new classification technique. This result adds additional interest to Grossberg's hypothesis that the mammalian cerebral cortex carries out local-in-time nearest matched filter classification of both auditory and visual sensory inputs as an initial step in sensory pattern recognition-which may help explain the almost instantaneous pattern recognition capabilities of animals.