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

Authors:
Yiannis Kokkinos;Konstantinos Margaritis
Affiliations:
Parallel and Distributed Processing Laboratory, Department of Applied Informatics, University of Macedonia, Thessaloniki, Greece;Parallel and Distributed Processing Laboratory, Department of Applied Informatics, University of Macedonia, Thessaloniki, Greece
Venue:
SETN'12 Proceedings of the 7th Hellenic conference on Artificial Intelligence: theories and applications
Year:
2012

Citing 9
Cited 0

Probabilistic neural networks

Neural Networks
Simulating Artificial Neural Networks on Parallel Architectures

Computer - Special issue: neural computing: companion issue to Spring 1996 IEEE Computational Science & Engineering
Weighted Parzen Windows for Pattern Classification

IEEE Transactions on Pattern Analysis and Machine Intelligence
Approximation algorithms for geometric problems

Approximation algorithms for NP-hard problems
Data Mining: Introductory and Advanced Topics

Data Mining: Introductory and Advanced Topics
Gap-Based Estimation: Choosing the Smoothing Parameters for Probabilistic and General Regression Neural Networks

Neural Computation
Novel Approaches to Probabilistic Neural Networks Through Bagging and Evolutionary Estimating of Prior Probabilities

Neural Processing Letters
Learning vector quantization for the probabilistic neural network

IEEE Transactions on Neural Networks
A neural network approach to statistical pattern classification by `semiparametric' estimation of probability density functions

IEEE Transactions on Neural Networks

Quantified Score

Hi-index	0.00

Visualization

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.