Analysis of MLP, CNN, and GNN Architectures for MNIST Digit Classification: Accuracy, Inference Time, and Memory Trade-Offs

This study evaluates the performance of three lightweight neural network architectures, Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and Graph Neural Network (GNN), on the MNIST digit classification task, using vector, grid, and graph-based representations, respectively. The research focuses on understanding the importance of model architecture for resource-constrained edge devices, where memory usage and inference speed are critical, and examining the tradeoffs between accuracy, inference time, and memory consumption according to preferred models. The findings show that CNN achieves the highest classification accuracy, GNN provides the lowest memory footprint, and MLP shows the lowest inference time. The results also underscore the trade-offs between model accuracy, inference time, and memory usage, providing insights for selecting the suitable architecture and data representations.