Spectral Regularization
Overview
Spectral regularization modifies learning by shrinking, filtering, or constraining directions identified through eigenvalues, singular values, or frequency components. Common approaches include ridge and Tikhonov filters, principal-component truncation, early stopping, spectral- or nuclear-norm penalties, and frequency-domain losses. These methods are studied in kernel learning and inverse problems as well as neural-network training, graph learning, and generative models, where they can improve generalization, stability, or representation quality. Theory provides adaptive and minimax guarantees in some settings, but a central challenge is choosing which directions to suppress: uniform shrinkage may discard useful structure, while selective control must avoid leaving unstable or collapsed modes unchecked.
Latest papers
47 papers3 papers in the last four weeks, down 50% on the four weeks before. 0.0% of all new papers.