Conceptual

Learned Robust Matrix Completion via Deep Unfolding

LRMC is a scalable non-convex algorithm for robust matrix completion (recovering a low-rank matrix from data that is both incomplete and outlier-corrupted) with proven linear convergence. Its free parameters are learned by deep unfolding — unrolling the iterative solver into a trainable network — and a feedforward-recurrent-mixed network extends unfolding from a fixed iteration count to effectively infinite iterations, beating prior methods on video background subtraction, ultrasound imaging, face modeling, and satellite cloud removal.