Training method that constrains weight matrices to lie on a fixed-norm hypersphere for improved stability and scaling.