Study details
Handwritten digit recognition remains useful for studying compact image-classification models, although MNIST no longer supports a state-of-the-art novelty claim. The study therefore treats MNIST as a controlled environment for examining accuracy–efficiency behavior in lightweight convolutional neural networks. It evaluates two five-layer CNN configurations across five parameter-variation cases, eight named optimizers plus a baseline control, five epoch settings, and augmented versus non-augmented data, producing 540 paired observations of accuracy and execution time. Peak accuracy and practical efficiency led to different choices: standard CNN-5LS with Nadam at 120 epochs after augmentation achieved 99.50% in 1,011 seconds, while RMSprop reached 99.47% at 15 epochs in 114 seconds. In the parameter-varied CNN-5L model, Case-1 was generally the most efficient and stable, while wider dilation and stride combinations often reduced reliability. Augmentation benefited weak or unstable settings most, and Ftrl did not learn meaningfully. The results support selecting lightweight CNN configurations through accuracy–time trade-offs within the bounded MNIST setting.