Application of Metaheuristic Algorithms for Determining the Structure of a Convolutional Neural Network with a Small Dataset

Inadequately labeled data can limit the accuracy of classification in image recognition tasks. Several methods have been proposed in the past to alleviate this limitation, such as transfer learning and data augmentation. However, the classification accuracy of the convolutional neural network (CNN) largely depends on its structural parameters, which are known as hyper-parameters. Therefore, in this paper, we introduce another method for minimizing the misclassification rate in a given small dataset by determining the hyper-parameters. The harmony search (HS) algorithm, improved harmony search (IHS) algorithm, self-adaptive global best harmony search (SGHS) algorithm, and novel global harmony search (NGHS) algorithm are applied for determining the optimal hyper-parameters. Additionally, we also compared the estimation performances of these four HS algorithms. It was finally observed that the HS and the IHS algorithms greatly outperform the other two algorithms.