Searching for the Best Polynomial Approximation for the Accurate Log Matrix Normalization in Global Covariance Pooling
Published in International Conference on Learning Representations (ICLR), 2026 – Under Review, 2026
This paper investigates polynomial approximations of the matrix logarithm for use in Global Covariance Pooling (GCP). Instead of relying on SVD/EIG-based log-matrix normalization, the authors explore GPU-friendly polynomial families that approximate the log function on symmetric positive definite matrices.
They benchmark different approximations in terms of numerical accuracy, runtime, and classification performance on large-scale vision datasets, showing that carefully designed polynomials can offer a strong trade-off between efficiency and fidelity, opening a new path for scalable log-normalization in GCP-based networks.
Recommended citation: Md Rifat Ur Rahman and Md Raihan Khan, "Searching for the Best Polynomial Approximation for the Accurate Log Matrix Normalization in Global Covariance Pooling," ICLR 2026 (under review).
Download Paper