Neural Networks with Differentiable Belief-Function Layers (arXiv:2610.30173)

Forum topic · 小凯 · 2026-10-24

Summary

This paper (arXiv:2610.30173, by Antonin Sulc) argues that Yager's 1987 evidence combination rule is the right choice for differentiable Dempster–Shafer layers in neural networks. Dempster's normalizer (1−κ)^{-1} introduces a (1−κ)^{-2} gradient term that diverges exactly when evidence sources conflict, whereas Yager's rule ⊞ is bilinear and gradient-stable. The authors build a deep evidential pipeline where every mass in the chain is a valid Dempster–Shafer mass function combined by Yager's rule; the output mass on the ignorance set m(Θ) acts as the network's "I don't know" signal. Trained only on hard labels and never exposed to annotator distributions, m(Θ) still ranks human disagreement across visual and linguistic modalities. Compared to Dempster's signal, Yager's vacuity is 9× larger on the most contested CIFAR-10H images and readable as ignorance without recalibration. It is competitive with strong single-pass baselines on OOD detection, while a clamped Dempster variant becomes gradient-unstable once the clamp is removed.

Paper Overview

Field: Machine Learning
Author: Antonin Sulc
Published: 2026-10-24
arXiv: 2610.30173

Why Yager's Rule Fits Backpropagation

Yager's 1987 rule of combination $\boxplus$ was proposed on epistemic grounds: under high conflict, mass should be redistributed to ignorance rather than renormalized onto the class simplex. The paper's key observation is that backpropagation needs the same rule:

Dempster's normalizer $(1{-}\kappa)^{-1}$ (where $\kappa$ is the conflict mass) puts a $(1{-}\kappa)^{-2}$ term in the gradient.
This term diverges exactly when evidence sources disagree.
Yager's $\boxplus$ is bilinear and has no such term, keeping gradients stable under conflict.

Deep Evidential Pipeline

The authors construct a pipeline in which every mass in the chain from heads to output is a valid Dempster–Shafer mass function, combined by Yager's rule. The output mass on the ignorance set, $m(\Theta)$ , serves as the network's "I do not know" signal.

Key Findings

Human disagreement ranking: The network is fit to hard labels and never sees an annotator distribution, yet $m(\Theta)$ orders human disagreement in two modalities — visual and linguistic.
Yager vs. Dempster: Dempster's signal also rises and ranks contested inputs about as well, but its normalizer divides the conflict out. Yager's vacuity is $9\times$ larger on the most contested CIFAR-10H images, so only Yager's is readable as an amount of ignorance without recalibration.
OOD detection: The method is competitive with strong single-pass baselines. A clamped Dempster variant scores higher but becomes gradient-unstable once the clamp is removed.

---

*Automatically collected on 2026-10-24. Original post: zhichai.net.*