Beyond ResNet: A Custom CNN Architecture for High-Performance CIFAR-10 Classification

This paper introduces a custom Convolutional Neural Network (CNN) architecture that achieves superior performance on the CIFAR-10 dataset compared to a standard ResNet-18 baseline. Through a methodical, step-by-step enhancement process, we demonstrate that a carefully designed and optimized classical CNN can outperform a more complex, modern architecture. My baseline model achieved a strong accuracy of 88.36%. I then systematically integrated and evaluated two key enhancements: channel attention mechanisms and stochastic depth regularization. The incorporation of stochastic depth proved particularly effective, yielding our best model which attained 89.90% accuracy; a significant improvement over the 80.55% accuracy achieved by a comparably tuned ResNet-18. This research challenges the automatic preference for very deep architectures with residual connections for standard benchmarks and underscores the potential of methodically refined custom designs.