Deep Unfolding
Overview
Deep unfolding studies how to convert iterative algorithms—such as proximal-gradient methods, ADMM, and conjugate-gradient solvers—into finite neural networks whose stages correspond to algorithmic updates. It matters when problems have known physical or mathematical structure, including image reconstruction, signal recovery, sensing, and large-scale optimization, where purely learned models may be data-hungry or hard to interpret. Common approaches learn update parameters and priors while preserving constraints or measurement models, sometimes adding adaptive steps, generative priors, or specialized modules for local and nonlocal features. A central challenge is balancing interpretability, speed, and robustness to new conditions against the flexibility and scalability of less structured learned models.
Latest papers
37 papers1 paper in the last four weeks, down 67% on the four weeks before. 0.0% of all new papers.