PaperlayerView exact revision on arXiv

Source-linked AI summary

Neural Networks with Differentiable Belief-Function Layers

Antonin Sulc

arXiv:2610.30173v1cs.AI

TL;DR

The paper asks how to implement belief-function layers in a back-propagated network without losing DST semantics or training stability. It builds a DST pipeline using Yager’s rule, whose output vacuity tracks human disagreement in visual and linguistic data and remains readable on its own scale.

  • Problem

    The paper asks whether belief-function layers can preserve DST semantics and training stability in back-propagated networks, where Dempster’s conflict normaliser destabilises optimisation at high conflict.

    Source: §1 · pp. 1–2

  • Method

    The MP-ENN uses a small focal set of singletons and Θ, with masses combined by Yager’s rule in a differentiable DST pipeline.

    Source: §1 · p. 2, §3 · p. 3

  • Results

    Yager’s vacuity rises with human disagreement in visual and linguistic data; on the most contested CIFAR-10H decile, it is 9× larger than Dempster’s (0.044 vs. 0.005).

    Source: §5.2 · pp. 6–7

  • Takeaways & Limitations

    The output m(Θ) is readable as an amount of ignorance on its own scale and tracks human judgments of “I do not know” without training on annotator distributions.

    Source: §6 · p. 8

    The focal set excludes masses on class subsets, Yager combination is not associative, and human-alignment numbers are single-seed.

    Source: §6 · p. 8

Abstract

from arXiv · show

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.

1 Introduction

The paper asks whether belief-function layers can preserve Dempster–Shafer semantics while remaining trainable by backpropagation. It argues that Yager’s rule addresses both the epistemic handling of conflict and the need for stable gradients.

1 Introduction

  • Dempster–Shafer theory represents ignorance with mass on Θ, unlike softmax classifiers that normalise mass across classes regardless of evidence.The paper asks whether this framework can be implemented in a back-propagated network without losing its semantics or training stability.Source: §1 · p. 1
  • Yager’s rule is bilinear and avoids Dempster’s conflict-dependent normaliser, which introduces a (1−κ)−2 gradient term that destabilises optimisation at high conflict.The paper presents this as the same rule required by epistemic treatment of conflict and by backpropagation.Source: §1 · p. 2
  • A small focal set F = {{ci}, Θ} keeps enumeration tractable and Bel/Pl exact, while learnable per-source discounting adapts reliability to inter-head agreement.The paper describes these design choices as removing remaining construction obstacles after adopting Yager’s rule.Source: §1 · p. 2
  • The paper contributes an end-to-end neural DST stack whose output vacuity separates absent from conflicting evidence, and tests whether m(Θ) tracks human visual and linguistic ambiguity.The network is trained on hard labels rather than annotator distributions.Source: §1 · p. 2

2 Related Work

Prior DST work has studied combination rules, while belief-function classifiers have generally aggregated evidence with Dempster’s rule. The paper situates its comparison of combination rules in back-propagated networks alongside single-pass and sampling-based uncertainty methods.

2 Related Work

  • The paper identifies its work as the first direct comparison of Yager and Dempster as combination layers in a back-propagated network.Earlier belief-function classifiers and evidential deep classifiers aggregate with Dempster’s rule.Source: §2 · p. 2
  • Single-pass methods and sampling alternatives summarize predictive uncertainty, mostly as a single scalar, while standard OOD baselines are collected in OpenOOD.The passage names SNGP, DUQ, EDL, Posterior Networks, MC Dropout, and Deep Ensembles among the methods discussed.Source: §2 · p. 2
  • Fig. 1 presents the MP-ENN forward pass, with each box implemented by the classical operator above it and forward and backward computation staying inside DST.The caption identifies the diagram’s scope but does not specify the individual operators.Source: Figure 1 · p. 3

3 Method

The method builds a quasi-Bayesian evidential network whose masses remain valid from evidence heads through fusion, using Yager’s rule to send conflict to ignorance. Learned head reliability and a hard-label objective complete the pipeline.

3.2 Mass-Primitive Evidence Heads

  • Evidence heads assign mass to class singletons and Θ on the quasi-Bayesian focal set, with W acting as the Dirichlet prior strength.The model uses F={{c1}, …, {cK}, Θ}, not a general belief function on 2^Θ.Source: §3 · p. 3, §3.2 · pp. 3–4

3.1 Overview of the Forward Pass

  • The forward pass composes mass assignment, α-discount, and Yager fusion, with output m(Θ) serving as vacuity, the OOD score, and the human-disagreement tracker.The chain rule passes through each layer unchanged.Source: §3.1 · p. 3

3.3 Yager Combination, and Why Not Dempster

  • Yager’s rule sends conflict to Θ rather than normalising it, while Dempster’s gradients grow without bound as conflict approaches one.On the focal set F, Yager’s largest partial derivative stays at 1.00 at every conflict level; Dempster’s reaches 4999.2 at κ=0.9998.Source: §3.3 · p. 4

3.4 Multi-Source Fusion with Learned Reliability

  • Later heads receive learned reliability discounts, then discounted masses are fused in a fixed left-fold order.Yager’s rule is commutative but not associative, so the output can depend on fusion order.Source: §3.4 · pp. 4–5

3.5 Training Objective

  • The hard-label objective rewards mass on the true class, penalizes vacuity, and uses an auxiliary term to prevent auxiliary heads from collapsing.The task term is proper given m(Θ), but the paper does not claim joint propriety over the full (K+1)-simplex.Source: §3.5 · p. 5

4 Theoretical Properties

The pipeline preserves valid mass functions and belief–plausibility identities by construction. Yager fusion is stable under differentiation, though its non-associativity motivates a fixed fusion order.

4 Theoretical Properties

  • Every per-source, discounted, and fused output remains a valid mass function with Bel ≤ Pl and exact belief–plausibility duality, but Yager fusion is not associative.The method therefore uses a fixed fusion order; it also implements averaging across all |S|! orderings.Source: §4 · p. 5
  • With one input mass fixed, Yager combination is linear and column-stochastic, making it 1-Lipschitz in ℓ1 and bounding each partial derivative by 1.The bound applies to every nonempty A and either input mass.Source: §4 · p. 5

5 Experiments

The experiments test whether output vacuity tracks human disagreement without annotator-distribution supervision, and assess its out-of-distribution detection performance. Yager vacuity tracks disagreement in visual and linguistic data, while Dempster ranks visual disagreement similarly but leaves much less readable vacuity.

5.2 Vacuity Tracks Human Disagreement (the Interpretability Test)

  • Yager’s vacuity rises with visual annotator entropy (ρ=0.278) and separates the most- from least-contested CIFAR-10H decile at AUROC 0.77.Across five entropy bins, vacuity increases from 0.008 to 0.043; Dempster’s is 0.044 versus 0.005 on the most-contested decile, despite similar ranking (ρ=0.263).Source: §5.2 · p. 6, Figure 2 · p. 7
  • On ChaosNLI, vacuity rises with human disagreement (ρ=0.346, p < 10^-40), reaching 0.076 on the most-contested quartile versus 0.036 on the least.Softmax entropy (ρ=0.390) and energy (ρ=0.346) match vacuity as scalar measures; vacuity also carries a [Bel, Pl] interval.Source: §5.2 · p. 7

5.3 OOD Detection (Secondary Capability)

  • Clamped Dempster scores highest among single-pass OOD methods, while Yager scores with energy; Dempster’s gradients are 1.52 versus Yager’s 0.21 at mean conflict 0.75.The Dempster comparison uses a clamp at 10^-8; without it, its gradients become unstable, while Yager needs no guard.Source: §5.3 · p. 7
  • Changing prior weight W from 1 to 100 keeps near-OOD AUROC in [0.855, 0.866], while changing λunc from 0 to 1 moves AUROC from 0.865 to 0.850.On CIFAR-100, fusion-order mass differences grow from 0.002 to 0.012 as mean conflict rises from 0.022 to 0.144, but vacuity AUROC varies by only 0.0004; inference takes 51.9 ms.Source: §5.3 · pp. 7–8

5.1 Setup

  • The OOD comparison evaluates ten methods across five seeds on CIFAR-10/OpenOOD, using macro AUROC and FPR95 averaged over seven OOD datasets.The setup includes CIFAR-100 as near-OOD and six far-OOD datasets; human-alignment estimates use a representative seed and neither annotator distribution is used for training.Source: §5.1 · p. 6, Table 1 · p. 8

6 Discussion and Conclusion

The paper argues that Yager’s rule both redirects conflict to ignorance and avoids Dempster’s gradient blow-up, enabling a differentiable evidential pipeline. Its output vacuity tracks human uncertainty, but the model’s focal-set scope and other limitations constrain interpretation.

6 Discussion and Conclusion

  • Yager’s rule assigns conflict to ignorance rather than normalising it, removing Dempster’s (1−κ)^-2 gradient blow-up from the combination step.The resulting pipeline combines Shafer’s mass and α-discount, Yager’s rule, and Jøsang’s vacuity reading in a single forward pass.Source: §6 · p. 8
  • The output m(Θ) is readable on its own scale and tracks human judgments of “I do not know” without training on annotator distributions.Source: §6 · p. 8
  • The focal set contains only singletons and Θ, excluding mass on subsets such as {ci, cj}; human-alignment estimates are single-seed.The learned α-discount is inactive rather than validated on the shared encoder, and Yager combination is not associative.Source: §6 · p. 8

Related papers

Loading 2610.30173v1…