Dynamic structural neural network

Artificial neural network (ANN) has been well applied in pattern recognition, classification and machine learning thanks to its high performance. Most ANNs are designed by a static structure whose weights are trained during a learning process by supervised or unsupervised methods. These training met...

Full description

Saved in:
Bibliographic Details
Published in:Journal of intelligent & fuzzy systems 2018-01, Vol.34 (4), p.2479-2490
Main Authors: Giap, Cu Nguyen, Son, Le Hoang, Chiclana, Francisco
Format: Article
Language:English
Subjects:
Artificial neural networks
Graph theory
Learning
Learning theory
Machine learning
Neural networks
Pathology
Pattern recognition
Training
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Artificial neural network (ANN) has been well applied in pattern recognition, classification and machine learning thanks to its high performance. Most ANNs are designed by a static structure whose weights are trained during a learning process by supervised or unsupervised methods. These training methods require a set of initial weights values, which are normally randomly generated, with different initial sets of weight values leading to different convergent ANNs for the same training set. Dealing with these drawbacks, a trend of dynamic ANN was invoked in the past year. However, they are either too complex or far from practical applications such as in the pathology predictor in binary multi-input multi-output (MIMO) problems, when the role of a symptom is considered as an agent, a pathology predictor’s outcome is formed by action of active agents while other agents’ activities seem to be ignored or have mirror effects. In this paper, we propose a new dynamic structural ANN for MIMO problems based on the dependency graph, which gives clear cause and result relationships between inputs and outputs. The new ANN has the dynamic structure of hidden layer as a directed graph showing the relation between input, hidden and output nodes. The properties of the new dynamic structural ANN are experienced with a pathology problem and its learning methods’ performances are compared on a real well known dataset. The result shows that both approaches for structural learning process improve the quality of ANNs during learning iteration.
ISSN:1064-1246
1875-8967
DOI:10.3233/JIFS-171947