Image Classification using Adaptive Multi-Module

For a classification using neural network, there exist many cases in which the distributions of classes are so complex that the classification with single network does not properly differentiate the given data into classes. This problem can be resolved if we employ multiple modules that can classify different data respectively. This paper proposes a new adaptive architecture for classification problems, and simulates its performance on image classification that is not easily classified using traditional neural network learning algorithms. Classification modules are added as learning proceeds dynamically, depending on the data. When a new module is introduced, it is trained for classification of the remaining data on which the currently existing module does not perform classification well. Simulation results show that the proposed Adaptive Multi-module Classification Network (AMCN) achieves accuracy improvement. It also outperforms single module classification case when they are compared in the condition of equal weight updates.