Huber-Wasserstein barycenters for robust distribution-valued data
arXiv:2608.13131
2026
Architecture
2 ideas extracted · analyzed Sep 1, 2026
What the math gives to ML
The paper introduces a robust Wasserstein barycenter by applying the Huber loss to each transport displacement before aggregation, rather than applying a robust loss to the final Wasserstein distance. This preserves quadratic sensitivity for small, informative displacements while clipping the influence of grossly corrupted regions, yielding a mean-to-median interpolation with a breakdown point near one half. The most transferable construction is a differentiable barycentric pooling or distribution-matching loss for neural modules whose objects are probability distributions, such as token histograms, uncertainty distributions, learned latent measures, or ensemble predictions. In one dimension, the method becomes a coordinatewise robust Huber location estimator on quantile functions, making an inexpensive GPU implementation possible.
Ideas from this paper
✓✓ Beats tuned baseline
2026
Replace ordinary Wasserstein or arithmetic pooling of distribution-valued features with a barycenter whose individual quantile displacements are Huberized. Small changes between input distributions remain averaged quadratically, while a corrupted token, expert, graph neighborhood, or augmentation cannot move the pooled distribution arbitrarily far. The module is especially cheap for one-dimensional distributions represented by fixed quantile vectors.
Useful7/10
Difficulty4/10
Novelty7/10
Unverified
2026
For models that predict probability distributions, replace the usual Wasserstein-2 loss or a Huber penalty on the final Wasserstein distance with a Huber penalty on quantile-by-quantile prediction errors. This suppresses gradients from localized outliers while retaining quadratic gradients on the majority of the distribution, which is useful for uncertainty prediction, histogram prediction, and distributional distillation.
Useful6/10
Difficulty3/10
Novelty5/10