Neural Networks with Differentiable Belief-Function Layers
Paper · https://arxiv.org/abs/2610.30173
arXiv:2610.30173v1 Announce Type: new Abstract: Yager's 1987 rule of combination $\boxplus$ was proposed on epistemic grounds: under high conflict, mass should be redistributed to ignorance rather than renormalised onto the class simplex. We observe that backpropagation needs the same rule: Dempster's normaliser $(1{-}\kappa)^{-1}$, where $\kappa$ is the conflict mass, puts a $(1{-}\kappa)^{-2}$ term in the gradient that diverges exactly when evidence sources disagree, while Yager's $\boxplus$ is bilinear and has no such term. Building on this, we construct a deep evidential 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. 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. Dempster's rises too and ranks contested inputs about as well, but its normaliser 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. For out-of-distribution detection it is competitive with strong single-pass baselines, while the clamped Dempster variant that scores higher becomes gradient-unstable once the clamp is removed.
First seen by Oriel on .
AI summary
The paper builds deep evidential networks using Yager’s combination rule, arguing it provides stable gradients under conflicting evidence and an interpretable measure of ignorance.
- Dempster’s normaliser introduces a gradient term proportional to $(1{-}\kappa)^{-2}$ that can diverge when evidence sources conflict; Yager’s rule is bilinear and avoids it.
- The pipeline combines valid Dempster–Shafer mass functions, with output mass on the ignorance set, $m(\Theta)$, used as an “I do not know” signal.
- Trained only on hard labels, the network’s ignorance score orders human disagreement in visual and linguistic modalities.
- On the most contested CIFAR-10H images, Yager’s vacuity is 9× larger than Dempster’s, whose normaliser divides conflict out.
- For out-of-distribution detection, the method is competitive with strong single-pass baselines; a higher-scoring clamped Dempster variant becomes gradient-unstable without the clamp.
#dempster-shafer theory #neural networks #uncertainty estimation #out-of-distribution detection
Generated by gpt-5.6-sol. Check the original for details.
Aliases
| Source | External ID | URL |
|---|---|---|
| arXiv AI Papers | oai:arXiv.org:2610.30173v1 | https://arxiv.org/abs/2610.30173 Read discussion |