We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
PULSATING MULTILAYER PERCEPTRON.
- Authors
PUPEZESCU, Valentin
- Abstract
The Knowledge Discovery in Databases represents the process of extracting useful information from data that are stored in real databases. The Knowledge Discovery in Databases process consists of multiple steps which include selection target data from raw data, pre-processing, data transformation, Data Mining and interpretation of mined data. As we see, the Data Mining is one step from the whole process and it will perform one of these Data mining task: classification, regression, clustering, association rules, summarization, dependency modelling, change and deviation detection. In these experiments we used one neural network (multilayer perceptron) that performs the classification task. This paper proposes a functioning model for the classical multilayer perceptron that is a sequential simulation of a Distributed Committee Machine. Committee Machines are a group of neural structures that work in a distributed manner as a group in order to obtain better classification results than individual neural networks. The classical backpropagation algorithm is modified in order to simulate the execution of multiple multilayer perceptrons that run in a sequential manner. The classification was made for three standard data sets: iris1, wine1 and conc1. In my case the backpropagation algorithm still consists of three well known stages: the feedforward of the input training pattern, the calculation of the associated output error, and the correction of the weights. The proposed model makes a twist for the classical backpropagation algorithm meaning that all the weights of the multilayer perceptron will be reset and randomly regenerated after a certain number of training epochs. This model will have a pulsating effect that will also prevent the blockage of the perceptron on poor local minimum points. This research is useful in the Knowledge Discovery in Databases process because the classification gets the same performance results as in the case of a Distributed Committee Machine.
- Subjects
ARTIFICIAL neural networks; DATA mining; KNOWLEDGE management; MULTILAYER perceptrons; ARTIFICIAL intelligence
- Publication
eLearning & Software for Education, 2016, Issue 1, p243
- ISSN
2066-026X
- Publication type
Article
- DOI
10.12753/2066-026X-16-035