Fast Gauss-Newton for Multiclass Cross-Entropy
Researchers have developed a Fast Gauss-Newton (FGN) method to approximate the generalized Gauss-Newton (GGN) curvature for multiclass cross-entropy. This new approach decomposes the standard GGN into a true-vs-rest term and a positive semidefinite within-competitor covariance term, dropping the latter to create an efficient under-approximation. The FGN method is exact for binary classification and can be solved efficiently using matrix-free conjugate gradient methods, showing promise for scaling up training. AI
IMPACT Introduces a more efficient approximation for training deep learning models with many classes, potentially speeding up convergence.