Invertible Convolution with Symmetric Paddings

Convolutional Neural Network (CNN) has achieved incredible success in computer vision. It is shown to be effective in building transformation between the highly complex distribution of image data domain to a latent feature domain distribution that is easy to conduct tasks like classification or generation. Recent studies have shown that there exist CNN operators that are invertible. This makes it possible to build bidirectional transformation between the image data domain and the latent feature domain. Such transformation is recognized to be valuable in unifying the formulation of generative and discriminative tasks and provides potential insights in revealing the interior process of CNNs.