Optimisation of convolutional neural network parameters using the bees algorithm

Packianather, Michael S. ORCID: https://orcid.org/0000-0002-9436-8206 and Alamri, Nawaf Mohammad H. 2024. Optimisation of convolutional neural network parameters using the bees algorithm. Pham, D. T. and Hartono, Natalia, eds. Intelligent Engineering Optimisation with the Bees Algorithm, Springer Series in Advanced Manufacturing, pp. 265-284. (10.1007/978-3-031-64936-3_13)

Full text not available from this repository.

Abstract

The convolutional neural network (CNN) is one of the most popular deep learning algorithms that deals mainly with image data. In this study, the application of the Bees Algorithm (BA), which behaves like honeybees, was used along with the Bayesian Optimisation (BO) approach to improve CNN performance (BA-BO-CNN). Applying the hybrid algorithm on Cifar10DataDir images increased the accuracy on the validation set from 80.72% for the existing BO-CNN to 82.22% for the hybrid algorithm. Applying BA-BO-CNN on digit datasets showed the same accuracy as the existing hybrid BO-CNN, but with a computational time shortened by 3 min and 12 s. Finally, concrete crack benchmark image data yielded almost similar results to existing algorithms. Similarly, a new hybrid Bees Convolutional Neural Network (BA-CNN) algorithm was proposed that uses BA to design better CNN topology by optimising its hyperparameters, which increases the network accuracy. Applying the hybrid algorithm to the Cifar10DataDir dataset yielded the same accuracy as existing CNNs and BO-CNNs. Applying it to the digits dataset produced the lowest computational time with 4 min and 14 s reductions compared to BO-CNN, so it is the best algorithm in terms of cost-effectiveness. Finally, applying it to concrete crack images produced similar results to the existing algorithms.

Item Type:	Book Section
Date Type:	Published Online
Status:	Published
Schools:	Engineering
ISBN:	978-3-031-64935-6
ISSN:	1860-5168
Last Modified:	10 Dec 2024 10:30
URI:	https://orca.cardiff.ac.uk/id/eprint/174613

Actions (repository staff only)

Edit Item