EVALUATING OPTIMAL CONFIGURATIONS OF SHALLOW CNNS ON LOW-COMPLEXITY DATASETS

Abstract

This study investigates the impact of network depth, kernel size configurations and convolutional layer types on the performance of convolutional neural networks (CNNs) for image classification tasks. The objective is to systematically assess how variations in kernel size, convolutional strategy, and depth contribute to classification accuracy and computational efficiency. We analyze both shallow and moderately deep network variants to disentangle the relative importance of kernel modifications versus depth scaling. Using the MNIST and CIFAR-10 datasets, we evaluate kernel sizes ranging from 3×3 to 11×11, convolutional configurations including standard, dilated, and strided convolutions, and network depths spanning from two to eight convolutional layers, across multiple input resolutions (32×32, 64×64, 128×128, and 256×256 pixels). Results indicate that for simple datasets like MNIST, depth scaling provides limited benefits, as high accuracy (~97–98%) is achievable even with shallow networks regardless of kernel variation. In contrast, for CIFAR-10, smaller kernels in early layers combined with increased depth capture fine-grained details more effectively, while larger kernels in later layers enhance generalization at higher resolutions. Dilated convolutions offer notable improvements for high-resolution images by reducing inference time (up to a 2× speedup at 256×256) without sacrificing accuracy, particularly in deeper models. Strided convolutions, although computationally efficient, consistently degrade accuracy due to spatial information loss across depths. These findings suggest that optimal performance arises from a balanced interplay between kernel configuration, convolutional strategy, and network depth. Future research may explore adaptive architectures that dynamically adjust these factors to improve robustness across diverse datasets and resolutions.