On the Selection of Hyperparameters in Convolutional Neural Networks
Donglin Wang, Qiang Wu · 2021 International Conference on Computational Science and Computational Intelligence (CSCI) · 2021
Convolutional neural networks (CNNs) have been showing great success in a variety learning tasks, especially object detection and image recognition. In the applications of CNN models, many hyperparameters, such as kernel size, channel size, learning rate, padding, and stride step, should be tuned and the success of CNN models highly depends on the correct selection of these parameters. However, it is very time-consuming and challenging to tune hyperparameters. In most cases one has to preselect some values for the hyperparameters based on experiences. In this paper we explore the impact of kernel size and channel size on the accuracy of feature extraction and imaging classification by CNN models and provide a practical guide to set up these hyperparameters. Particularly for kernel size we found the choice could be quite flexible for an easy task while the best choice should be in a range for a difficult task, and the range shrinks as the feature to background ratio becomes smaller.