A mutual information measure for feature selection with application to pulse classification

This paper presents a method of constructing a low dimensional representation of a pulse signal that preserves the information necessary to classify the pulse. A large set of “atomic features” are generated from dilations and shiftings of a family of one or more mother wavelet functions. Each atomic feature is used to generate a feature vector element by taking the inner product of the pulse with the atomic feature. The “best” feature vector elements are selected according to the mutual information between the pulse category and feature vector element values over all pulse realizations. The result is a low dimensional transformation that retains the information necessary to discriminate one category from another. This paper presents an algorithm for computing the above mutual information measure on a CNAPS, a 512 processor SIMD parallel computer