Analysing the empirical consequences of hidden layer configurations

This paper explores the impact of structural variations in convolutional neural networks (CNNs) by comparing three distinct architectural configurations: parallel, converging, and diverging hidden-layer designs. Each configuration is evaluated on three standard grayscale image classification datasets-MNIST, Fashion-MNIST, and a grayscale-converted version of CIFAR-10-while maintaining identical parameter budgets and training conditions. Performance is assessed using F1 Score, Precision, Recall, training time, and a custom metric, Signed Confidence Average (SCA), to provide deeper insight into prediction confidence and stability. Empirical results show that parallel architectures generally achieve the highest classification accuracy, while diverging architectures demonstrate faster and smoother convergence with only marginal reductions in performance. In contrast, converging architectures consistently underperform across all metrics and datasets. These findings suggest that structural design plays a significant role in learning efficiency and representation quality, with diverging models offering a compelling trade-off between speed and performance. The study underscores the importance of architectural configuration in CNN design and opens new directions for lightweight, efficient model development.