Parallelism, localization and chain gradient tuning combinations for fast scalable probabilistic neural networks in data mining applications

Abstract

This work investigates the scalability of Probabilistic Neural Networks via parallelization and localization, and a chain gradient tuning. Since PNN model is inherently parallel three common parallel approaches are studied here, namely data parallel, neuron parallel and pipelining. Localization methods via clustering algorithms are utilized to reduce the hidden layer size of PNNs. A problem of localization may be present in the case of multi-class data. In this paper we propose two simple fast approximate solutions. The first is using sigma smoothing parameters obtained from the parallel PNN initial training directly to clustering. In this case a substantial reduction of neurons is achieved without significant loss of recognition accuracy. The second is an effort for an additional tuning. Via confidence outputs we employ a chain training approach to tune for the best possible PNN architecture. © 2012 Springer-Verlag.