Parallel and local learning for fast probabilistic neural networks in scalable data mining

Authors:
Yiannis Kokkinos;Konstantinos G. Margaritis
Affiliations:
University of Macedonia, Thessaloniki, Greece;University of Macedonia, Thessaloniki, Greece
Venue:
Proceedings of the 6th Balkan Conference in Informatics
Year:
2013

Citing 11
Cited 0

Probabilistic neural networks

Neural Networks
Local learning algorithms

Neural Computation
Local algorithms for pattern recognition and dependencies estimation

Neural Computation
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
Parallel Implementation of Back-Propagation Algorithm in Networks of Workstations

IEEE Transactions on Parallel and Distributed Systems
Novel Approaches to Probabilistic Neural Networks Through Bagging and Evolutionary Estimating of Prior Probabilities

Neural Processing Letters
Fast and Scalable Local Kernel Machines

The Journal of Machine Learning Research
Probabilistic neural-network structure determination for pattern classification

IEEE Transactions on Neural Networks
Learning vector quantization for the probabilistic neural network

IEEE Transactions on Neural Networks
Maximum likelihood training of probabilistic neural networks

IEEE Transactions on Neural Networks

Quantified Score

Hi-index	0.00

Visualization

Abstract

Probabilistic Neural Networks (PNN) represent knowledge in the form of simple well understood Bayesian models, which are most suitable for data mining applications that also need confidence levels. However if the number N of pattern neurons grows large, the evaluation of a single unknown sample has still O(N) complexity, which is not scale well. Parallelism can efficiently work towards speeding up the large neural networks. Whereas the PNN run time can be reduced by parallelism, its computational cost can be decreased by keeping only the globally most important pattern neurons. While such parsimonious global models have been studied for a long time, there is another way that is related to local learning algorithms. Maintain all N pattern neurons but to classify any unknown x use only the local ones, the k nearest neighbor to x neurons. This local learning PNN is investigated here. By using confidence ratio outputs we optimize the number of k nearest neighbor neurons for the best PNN performance. Combined with the parallel approach using p processors the method can reduce the cost of classifying an unknown x from O(N) to O(k/p).