Abstract
This research explores the trade-off between classification accuracy and computational efficiency in deep learning by conducting an empirical comparison of compact, optimized versions of the Visual Geometry Group (VGG) and Residual Neural Network (ResNet) architectures using the CIFAR-10 dataset. The main objective was to identify which design approach VGG’s straightforward sequential structure or ResNet’s skip-connection-based parameter efficiency achieves a more effective balance in resource-limited settings. The results show that while ResNet is more parameter-efficient, requiring 1.23 million parameters compared to VGG’s 5.35 million, it exhibited slower inference performance (6.71 ms per image versus 6.04 ms for VGG). Moreover, ResNet demonstrated weaker robustness to noisy inputs, with its accuracy decreasing to 35.28%, whereas VGG’s accuracy declined more moderately to 48.16%. Overall, the study highlights that in low-resolution image classification tasks, simpler architectures can outperform more complex models that are theoretically more efficient. These results offer valuable guidance for practitioners seeking to select and optimize convolutional neural networks under strict computational constraints.