Projection Pursuit CPCANet for Domain Generalization

Researchers introduced PP-CPCANet, a covariance-free domain generalization method for robust representation learning.

It replaces batch covariance estimation with a global orthogonal basis on the Stiefel manifold and a detached-median dispersion objective.

The approach learns common principal components more stably, helping models build domain-invariant features under distribution shift.

Experiments on four benchmarks showed better performance and more stable optimization than earlier CPCANet-style methods.