Comparative analysis of VGG, ResNet, and GoogLeNet architectures evaluating performance, computational efficiency, and convergence rates

Abstract

This paper conducts an in-depth comparative analysis of three foundational machine learning architectures: VGG, ResNet, and GoogLeNet. The focus of the evaluation is their performance metrics on the CIFAR-100 dataset, a widely adopted benchmark in the field. Employing a comprehensive set of evaluation metrics, this investigation assesses not only testing accuracy but also the rate of training convergence and computational efficiency, providing a holistic perspective on the architectures' capabilities. Through rigorous experimentation, we elucidate the inherent advantages and drawbacks associated with each of these architectures. For instance, our findings delve into the nuances of how different architectures fare in terms of computational resources, which is vital for deployment in resource-constrained environments. Additionally, this study extends the analysis to explore the effect of hyperparameter settings, particularly learning rates, and the utility of data augmentation techniques in modulating the overall performance of each architecture. The ultimate objective is to furnish empirical insights that will assist researchers and practitioners in making well-informed choices when selecting a machine learning architecture for their specific application requirements.