Fast and High-performance Multi-convolution Deep Neural Network Structure with Residuals
Yunpeng Wang, Yasutaka Fujimoto · 2018
Very deep convolution neural networks show a great improvement over competitive benchmarks. But the depth also brings extremely high computational cost. In this paper, inspired by Inception module, we introduce a new convolutional neural network module that combines residual structure and multiple convolution. A residual structure is mainly adopted to solve the gradient vanishing problem. And unlike the concatenation in the inception module, multiple convolution is used to find the most proper feature maps through self-optimizing training. With this module, we no longer need to carefully optimize the convolution structure of network, but can attain state-of-the-art results on CIFAR-10, MNIST and CIFAR-100 with 26 layers and only 19k parameters.