arXiv · 2611.19065
A Fast Algorithm for the Familywise Loss
Abstract
The familywise (FW) loss for multiclass classification, introduced by Duchi, Ruan, and Khosravi, achieves universal equivalence to the 0-1 loss, and thus has stronger theoretical properties than cross entropy (CE). Prior work, however, left open whether FW is as computationally efficient as CE. First, on the theoretical side, we prove that the FW gradient can be computed in $O(K)$ time, where $K$ is the number of classes. Second, on the practical side, we develop optimized kernels for computing the forward and backward passes of FW which outperform CE in terms of speed on both CPU and GPU.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Leelavathi Raja, Amirreza Eshraghi, Vinay Prabhakar, Mark Blacher, Han Bao, Yutong Wang. 2026-11-14. A Fast Algorithm for the Familywise Loss. https://arxiv.science/abs/2611.19065
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.