<ul><li>The singular values of convolutional mappings contain valuable spectral properties that can enhance the generalization and robustness of convolutional neural networks.</li><li>Computing singular values is usually resource-intensive, especially for large matrices representing convolutional mappings with high-dimensional inputs and many channels.</li><li>This work introduces a novel approach based on local Fourier analysis to efficiently compute singular values of convolutional mappings with a complexity of O(N).</li><li>The proposed method is scalable and provides a practical solution for calculating the complete set of singular values and corresponding singular vectors for high-dimensional convolutional mappings.</li></ul>

LFA applied to CNNs: Efficient Singular Value Decomposition of Convolutional Mappings by Local Fourier Analysis

Discover more