Neural PCA and Maximum Likelihood Hebbian Learning on the GPU

Abstract

This study introduces a novel fine-grained parallel implementation of a neural principal component analysis (neural PCA) variant and the maximum Likelihood Hebbian Learning (MLHL) network designed for modern many-core graphics processing units (GPUs). The parallel implementation as well as the computational experiments conducted in order to evaluate the speedup achieved by the GPU are presented and discussed. The evaluation was done on a well-known artificial data set, the 2D bars data set.

Soft Computing Techniques for Skills Assessment of Highly Qualified Personnel

“…AI has been successfully applied to many different fields as, for example, feature selection [18,19]. In this study, an extension of a neural PCA version [20,21] and further EPP [22,23] extensions are used to select the most relevant input features in the data set and to study their internal structure. Some projection methods such as PCA [20,21], MLHL [22] and CMLHL [24][25][26] are applied to analyse the internal structure of the data and find out, which variables determine this internal structure and what they affect in this internal structure.…”

“…Interestingness is usually defined in terms of how far the distribution is from the Gaussian distribution [27]. One neural implementation of EPP is MLHL [22]. It identifies interestingness by maximizing the probability of the residuals under specific probability density functions that are non-Gaussian.…”

“…An extended version of this model is the CMLHL [24,28] model. CMLHL is based on MLHL [22] adding lateral connections [24,28], which have been derived from the Rectified Gaussian Distribution [27]. The resultant net can find the independent factors of a data set but does so in a way that captures some type of global ordering in the data set.…”

“…Several Soft Computing techniques have been used previously in different areas [6,7] for managing large amounts of high dimensional information. Specifically, within the various existing AI techniques, "Exploratory Projection Pursuit" (EPP) algorithms [22][23][24], allow the visualization of high dimensional information in a compressible way for the user [8][9][10].…”

“…In particular, three EPP algorithms were applied in this study: "Principal Components Analysis" (PCA) [20,21], "Maximum Likelihood Hebbian Learning" (MLHL) [22] and "Cooperative Maximum Likelihood Hebbian Learning" (CMLHL) [24,28]. These algorithms were employed to a novel and real dataset based on the assessments of university students' skills referring to the field of odontology.…”