Network Architecture Design: Large Kernel Is What You Need
Shuchang Ye, Qinyi Cao · 2023
We revisited the principle of neural network architecture design and its bottlenecks. Since VGG, the number of parameters is exploding. People scarify the size and efficiency of neural networks for little increase in accuracy, which means the contribution per parameter is getting lower and lower. This study aims to seek a way to design neural networks wisely. Inspired by the achievements of Vision Transformer, we infer that the size of the Effective Receptive Field can significantly affect the performance and efficiency of neural networks. Based on RepLKNet's algorithms in accelerating large kernel computing and the structure of residual-assisted large kernel design, a method to break the bottleneck of convolutional neural networks is proposed.