Invertible Network

Invertible networks are neural network architectures designed to learn bijective mappings between input and output data, enabling both forward and inverse transformations. Current research focuses on applying these networks to diverse problems, including image processing (e.g., compressed sensing, denoising, rescaling), medical imaging (e.g., PET attenuation correction, CT generation), and physical system modeling (e.g., hemodynamic model inversion), often employing variations of normalizing flows or incorporating techniques like optimal transport for improved performance and interpretability. The ability to perform exact inversions offers significant advantages in terms of memory efficiency, improved model understanding, and the potential for novel applications in fields requiring both forward and inverse modeling.