PP-CPCANet addresses rank-deficient covariance estimation in mini-batch training by learning a global orthogonal basis on the Stiefel manifold using the Cayley transform. A symmetry-breaking detached-median dispersion objective is introduced to extract robust common principal components.
HOW THIS AFFECTS YOU
●
researcherYou can leverage this covariance-free framework to improve representation robustness against distribution shifts during mini-batch training.