Optimal evolutionary framework-based activation function for image classification
View/ Open
Neagu_et_al_Knowledge_Based_Systems (1.424Mb)
Download
Publication date
2024-09Keyword
Machine learningDecision support
Image classification
Genetic algorithm
Activation functions
Neural networks
Rights
© 2024 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).Peer-Reviewed
YesOpen Access status
openAccessAccepted for publication
2024-05-29
Metadata
Show full item recordAbstract
Typically, supervised Machine Learning (ML)-based image classifiers leverage algorithms derived from either Artificial Neural Networks (ANNs) or optimal separating hyperplane (OSH)-based algorithms. However, despite recent progress has been made to enhance ANNs’ classification performance via the Rectified Linear Unit (ReLU)-based activation functions (AFs), there is currently no AF that scales across and benefit both ANNs and OSH-based classifiers. Moreover, the lack of globally optimal AFs leads to a high variance in image classification-related results. Thus, this study seeks to overcome this limitation by implementing a next-generation evolutionary framework (‘ActiGen’) to generate a novel and more reliable AF, which can scale to two families of AFs for two classifiers. The proposed evolutionary knowledge-based framework leverages a Multi-Objective (MO) optimisation method based on Genetic Algorithms (GA), or ‘MOGA’, to improve the generalisation of such classifiers. This evolutionary framework and its generated AF are validated using nine open-access datasets: seven image-based datasets, consisting of 22,136 images in total, and two large (561 features for 10,929 instances, 124 features for 1,700 instances) tabular datasets. These diverse datasets include both binary and multi-class classification, such as images of breast masses, those acquired via cardiac computed tomography, photos of famous people from the Internet, images of handwritten digits and those drawn on a graphics tablet, human faces with different lighting, details, and expressions, smartphone-related data captured during various activities and postural transitions, and clinical data on complications of myocardial infarction. Findings demonstrate that the proposed evolutionary optimisation framework (‘ActiGen-MOGA’) was able to generate a novel scalable AF, which led to achieve the highest classification performance and the fastest convergence across six out of nine datasets. In the best classification task, the ActiGen-MOGA-based AF led to a classification performance of 80 % and 78 % higher than the polynomial and Rectified Linear Unit (ReLU) AFs respectively.Version
Published versionCitation
Parisi L, Neagu CD, RaviChandron N et al (2024) Optimal evolutionary framework-based activation function for image classification. Knowledge-Based Systems. 299: 112025.Link to Version of Record
https://doi.org/10.1016/j.knosys.2024.112025Type
Articleae974a485f413a2113503eed53cd6c53
https://doi.org/10.1016/j.knosys.2024.112025