This paper investigates the accuracy of Convolutional Neural Networks (CNNs) under constrained time cost, a critical issue in industrial and commercial applications. The authors explore the trade-offs between network depth, width, filter sizes, and strides to design efficient models while maintaining or improving accuracy. Through a series of controlled experiments, they modify a baseline model to preserve its time complexity while achieving competitive accuracy on the ImageNet dataset. The resulting model achieves a top-5 error rate of 11.8% with a 20% faster training speed compared to "AlexNet," demonstrating the effectiveness of their approach. The paper also discusses the importance of depth in improving accuracy and the limitations of increasing depth without proper adjustments. Additionally, it introduces techniques such as delayed subsampling of pooling layers and layer replacement to further optimize the models. The findings provide insights into designing CNNs that balance accuracy and computational efficiency, making them suitable for real-world applications.This paper investigates the accuracy of Convolutional Neural Networks (CNNs) under constrained time cost, a critical issue in industrial and commercial applications. The authors explore the trade-offs between network depth, width, filter sizes, and strides to design efficient models while maintaining or improving accuracy. Through a series of controlled experiments, they modify a baseline model to preserve its time complexity while achieving competitive accuracy on the ImageNet dataset. The resulting model achieves a top-5 error rate of 11.8% with a 20% faster training speed compared to "AlexNet," demonstrating the effectiveness of their approach. The paper also discusses the importance of depth in improving accuracy and the limitations of increasing depth without proper adjustments. Additionally, it introduces techniques such as delayed subsampling of pooling layers and layer replacement to further optimize the models. The findings provide insights into designing CNNs that balance accuracy and computational efficiency, making them suitable for real-world applications.