LAUSR

Convolution networks continue to create state-of-the-art results in computer vision, and the Residual Network is an important milestone. In the original residual network, 1 $$\times $$× 1 convolution with stride 2 is used as the projection to do the linear transformation between feature maps of different sizes and different number of channels. This projection structure does not satisfy the concept of residual learning and is not able to use all of the input information. We propose a method which will make the Residual Network completely free of this structure and realize what shortcut connections should be. Compared with the original Residual Network, our models achieve higher learning efficiency and higher average performance with fewer parameters and lower computational complexity on CIFAR-10/100.