1. Optimal transport is a useful tool for designing unsupervised deep learning networks for inverse problems like denoising, deconvolution, and accelerated MRI. 2. CycleGAN can be interpreted as minimizing two Wasserstein distances between the input and target spaces in an unsupervised manner. 3. The document presents applications of CycleGAN and its variations to multiphase CT denoising, microscopy deconvolution, and accelerated MRI, showing it can produce good results without supervised labels.