Classification Performance Analysis of Weight Update Method Applied to Various ConvNet Models

Abstract

Research on the artificial intelligence is increasing with the improvement of computing power and the development of algorithm theory. In particular, the deep neural network, which is a field of machine learning, is widely used in artificial intelligence because it can process data that cannot be solved by conventional shallow neural networks more effectively. Implementation of a deep neural network is generally based on popularized neural networks with excellent generalization performance, which saves time and effort. However, it is difficult to guess which deep neural networks and optimization methods can achieve the best performance in their dataset. In this paper, we have tested the four convolutional neural networks (ConvNet) and four weight update methods. Experiments were conducted using a 5-fold cross-validation based on insect image dataset. As a result, the ResNet-50 and AdaDelta combination showed the best performance (89.98 ± 1.40)% in the insect dataset.