PP-CPCANet removes covariance bottlenecks and reports SOTA on four DG benchmarks

unimelb-nlp · hf · 2026-07-29

PP-CPCANet is a new domain generalization method that removes the covariance-estimation bottleneck in CPCANet.

The paper proposes a covariance-free framework that learns a global orthogonal basis on the Stiefel manifold and optimizes it jointly with network parameters via the Cayley transform. It also introduces a detached-median projection pursuit dispersion objective to improve robustness and optimization stability. On four domain generalization benchmarks, the authors report state-of-the-art results with stable training.

Original post →

More from Research

Research channel →