Scale Invariant Probabilistic Neural Networks
Romin Ebrahimi · Texas ScholarWorks (Texas Digital Library) · 2018
This paper introduces a scale invariant version of the original PNN proposed by Specht [1] with the added functionality of allowing for smoothing along multiple dimensions. By using a general multivariate gaussian kernel for density estimation, the pattern units are scale invariant while accounting for covariances within the data set. Additionally, we detail an optimization procedure for selecting elements of a smoothing matrix. Finally, a condensed version of the Scale Invariant PNN (SPNN) is proposed, which can be used for latent class analysis. An implementation of the SPNN is written in the R statistical programming language and is available on CRAN under the package name spnn.