Orthogonality Constraint

A training technique that imposes row-wise orthogonal or block-orthogonal constraints on neural network weight matrices, aiming to improve optimization geometry and representation quality; it appears in research reports, commonly in fine-tuning settings.