Depth vs. Complexity: A Comparative Study of Neural Network Architectures in Image Classification

Abstract

There is a growing requirement for image classification algorithms in a plethora of fields, including medical imaging, autonomous vehicles, surveillance, etc. To streamline the process of designing such algorithms to accomplish such a task, one must be aware of the strengths and drawbacks of existing models. This paper investigates the performance of various image classification algorithms, focusing on the dynamic between model depth and complexity, and their effect on accuracy. This study utilizes three datasets - MNIST, Fashion MNIST, and CIFAR10 - to conduct a comprehensive analysis of six distinct image classification architectures. There is a discernible accuracy gradient as one traverses model complexities, from the standard Multilayer Perceptrons (MLPs) to a Visual Transformer (ViT). Training a ViT requires large amounts of computational resources, yet the investment is justified by the remarkable accuracy it achieves.  However, it is always more efficient to use a model that fits the scale of the data. No model is the best for every dataset, and data complexity plays a vital role in determining the optimal model architecture for any data.