Dynamic construction of multilayer neural networks for classification

Authors:
Jiqian Liu;Yunde Jia
Affiliations:
Beijing Laboratory of Intelligent Information Technology, School of Computer Science, Beijing Institute of Technology, Beijing, P.R. China;Beijing Laboratory of Intelligent Information Technology, School of Computer Science, Beijing Institute of Technology, Beijing, P.R. China
Venue:
ISNN'11 Proceedings of the 8th international conference on Advances in neural networks - Volume Part I
Year:
2011

Citing 6
Cited 0

A theory of the learnable

Communications of the ACM
Multilayer feedforward networks are universal approximators

Neural Networks
Neural networks and the bias/variance dilemma

Neural Computation
Support-Vector Networks

Machine Learning
The strength of weak learnability

SFCS '89 Proceedings of the 30th Annual Symposium on Foundations of Computer Science
The sample complexity of pattern classification with neural networks: the size of the weights is more important than the size of the network

IEEE Transactions on Information Theory

Quantified Score

Hi-index	0.00

Visualization

Abstract

There are several drawbacks of multilayer neural networks (MLNNs) including the difficulty of determining the number of hidden nodes and their black box nature. We propose a new dynamic construction mechanism for MLNNs to overcome such inherent drawbacks. The main goal of our work is to train a hidden neuron and assemble it to the network dynamically while making the learning error smaller and smaller. In this paper, a hidden neuron carries out the function of a linear classifier which answers yes(Y) or no(N) to whether the input data belongs to the specific class. We call such a linear classifier a Y/N classifier and call the hidden neuron a Y/N neuron. The number of Y/N neurons are determined self-adaptively according to the given learning error and then successfully avoid the overlearning problem. The dynamically constructed MLNN with Y/N neurons is called a Y/N neural network. We prove that a Y/N neural network can always converge to the required solution and illustrate that Y/N neural networks can be applied to very complex classification problems.