Comparitive analysis of Alexnet, GoogLeNet and EffecientNet using CIFAR-100 dataset

The CIFAR-100 dataset was implemented to help conduct this effective comparison study across the following mod- els—AlexNet [8], GoogLeNet [8], and EfficientNetB0 [7]—un- der designs of animal, human, and many other classifications, respectively. To prove that the models differ significantly with respect to the time of training, accuracy, and loss, we ran our tests. Although EfficientNetB0 takes the longest to train, at 47 minutes per epoch, it obtained the lowest final loss and best validation accuracy of 92.3%. Validation accuracy of 66% and a training period of 10 minutes were the well-balanced with GoogLeNet. AlexNet was the fastest with its 3-minute training session, but then the least accurate at 35% together with the most loss. Accuracy is the main factor for which EfficientNetB0 should be used. GoogLeNet should be used where accuracy and time to train need to be balanced, and AlexNet should be the model to use if training more quickly is imperative, even if accuracy suffers as a result. Future research may then look into performance improvement and ways to apply these models. It may include data augmentation, transfer learning, advanced optimization approaches, hybrid approaches, validation on more real-world datasets beyond this bench marking dataset, and deployment in edge computing. Index Terms—AlexNet, Convolutional neural networks, CIFAR-100 dataset, Deep neural networks, EffecientNetB0, GoogLeNet, Image processing.