An Integrative Neural Network with Feedback Control for Classification

This paper presents a novel integrative neural network, which contains a feedback loop for adaptive control of learning. Instead of designing a single classifier for the classification task, a finite number of classifiers are simultaneously applied and all outputs from the individual classifiers are processed by the integrative neural network. The stability conditions and supervised learning algorithms are derived for the artificial neural network. We have applied it to unconstrained handwritten numbers recognition. Experiments are carried out and the results are compared to that of multi-layer perceptron. They show that the proposed integrative neural network with feedback control has a better classification rate with no decrease of reliability. This type of neural network scheme provides an alternative approach for ensemble learning.