Invert to Learn to Invert

Abstract

Iterative learning to infer approaches have become popular solvers for inverse problems. However, their memory requirements during training grow linearly with model depth, limiting in practice model expressiveness. In this work, we propose an iterative inverse model with constant memory that relies on invertible networks to avoid storing intermediate activations. As a result, the proposed approach allows us to train models with 400 layers on 3D volumes in an MRI image reconstruction task. In experiments on a public data set, we demonstrate that these deeper, and thus more expressive, networks perform state-of-the-art image reconstruction.

Cited by in corpus (14)

• Deep Learning in Photoacoustic Tomography: Current approaches and future directions
• RE-MIMO: Recurrent and Permutation Equivariant Neural MIMO Detection
• Measuring Robustness in Deep Learning Based Compressive Sensing
• Invertible generative models for inverse problems: mitigating representation error and dataset bias
• Deep Unfolding with Normalizing Flow Priors for Inverse Problems
• Preconditioned training of normalizing flows for variational inference in inverse problems
• $\Sigma$-net: Systematic Evaluation of Iterative Deep Neural Networks for Fast Parallel MR Image Reconstruction
• Understanding and Mitigating Exploding Inverses in Invertible Neural Networks
• iUNets: Fully invertible U-Nets with Learnable Up- and Downsampling
• i-RIM applied to the fastMRI challenge
• MRI Image Reconstruction via Learning Optimization Using Neural ODEs
• Experimental design for MRI by greedy policy search
• Over-and-Under Complete Convolutional RNN for MRI Reconstruction
• Sum-Product-Transform Networks: Exploiting Symmetries using Invertible Transformations