Learning to solve inverse problems using Wasserstein loss
arXiv:1710.10898
Abstract
We propose using the Wasserstein loss for training in inverse problems. In particular, we consider a learned primal-dual reconstruction scheme for ill-posed inverse problems using the Wasserstein distance as loss function in the learning. This is motivated by miss-alignments in training data, which when using standard mean squared error loss could severely degrade reconstruction quality. We prove that training with the Wasserstein loss gives a reconstruction operator that correctly compensates for miss-alignments in certain cases, whereas training with the mean squared error gives a smeared reconstruction. Moreover, we demonstrate these effects by training a reconstruction algorithm using both mean squared error and optimal transport loss for a problem in computerized tomography.
11 pages, 3 figures
References in corpus (5)
- Adam: A Method for Stochastic Optimization
- TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems
- SGDR: Stochastic Gradient Descent with Warm Restarts
- Deep Generative Adversarial Networks for Compressed Sensing Automates MRI
- Recurrent Inference Machines for Solving Inverse Problems
Cited by in corpus (10)
- Deep Bayesian Inversion
- Recurrent Generative Adversarial Networks for Proximal Learning and Automated Compressive Image Recovery
- Task adapted reconstruction for inverse problems
- Interpolation and Extrapolation of Toeplitz Matrices via Optimal Mass Transport
- Big in Japan: Regularizing networks for solving inverse problems
- Learning step sizes for unfolded sparse coding
- Multi-Marginal Optimal Mass Transport with Partial Information
- Learning the geometry of wave-based imaging
- Unpaired Single-Image Depth Synthesis with cycle-consistent Wasserstein GANs
- Inverse Problem of Nonlinear Schrödinger Equation as Learning of Convolutional Neural Network