Accuracy–Efficiency Analysis of Optimizer Variants and Convolutional Parameter Configurations for Handwritten Digit Recognition

Abstract

Handwritten digit recognition remains useful for studying compact image-classification models, although the MNIST (Modified National Institute of Standards and Technology) benchmark is no longer sufficient for claiming state-of-the-art novelty. Accordingly, this study frames MNIST as a controlled setting for analyzing accuracy-efficiency behavior in lightweight convolutional neural networks (CNNs). It examines how optimizer choice, epoch count, data augmentation, and convolutional parameter settings jointly affect classification accuracy and execution time. The experimental design covers two 5-layer CNN configurations, five parameter-varying cases, eight named optimizer settings, one baseline control setting, five epoch values, and two augmentation conditions. In total, 540 paired observations of accuracy and execution time are analyzed. The results show that peak accuracy and practical efficiency lead to different conclusions. The best observed accuracy was 99.50% for the standard CNN-5LS configuration with Nadam at 120 epochs after augmentation, but this required 1,011 s. RMSprop reached 99.47% after augmentation at both 15 and 120 epochs in the standard configuration, while the 15-epoch setting required only 114 s. In the parameter-varied CNN-5L model, Case-1 was generally the most efficient and stable configuration, while wider dilation and stride combinations frequently reduced reliability. Data augmentation improved weak or unstable settings more strongly than already stable configurations, and Ftrl did not produce meaningful learning in these experiments. These findings suggest that lightweight CNN design for handwritten digit recognition should be evaluated through accuracy-time trade-offs, not only through maximum accuracy. The scope is practical configuration guidance within the MNIST experimental setting.