acceptodds
Withdrawn from ICLR 2027

Dynamic Negation of Basic Belief Assignments in Dempster-Shafer Theory with Applications in Classification and Low-Light Image Enhancement

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The effective modelling of uncertain information is the key to improving the level of decision-making. Dempster-Shafer theory (DST) can express unknown and partial reliability explicitly. However, existing methods based on DST focus mostly on the accumulation of positive evidence and ignore the counterbalancing effect of negative evidence on the belief distribution, which leads to one-sidedness of decision-making in scenarios of intensified conflict. To this end, this paper proposes a dynamic negation of basic belief assignment (DNBBA), which measures the strength of refutation. In addition, we analyse the properties of DNBBA. On this basis, a decision-making network based on DNBBA (DNBBA-DNet) is proposed, which achieves the migration of the decision paradigm from positive accumulation to dialectical checks and balances. The experimental results demonstrate that the proposed method outperforms state-of-the-art methods on benchmark datasets, enhancing the classification accuracy and overall performance of low-light image enhancement (LLIE) tasks.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Dempster-Shafer theory, basic belief assignment, classification, low-light image enhancement

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_KZXPOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors propose a method involving the dynamic negation of basic belief assignment and apply it to low-illumination scenarios. With all due respect, I find the description of the method, the underlying motivation, and the experimental setup to be unclear.

Critical strengths

The authors propose a method involving the dynamic negation of basic belief assignment and apply it to low-illumination scenarios. With all due respect, I find the description of the method, the underlying motivation, and the experimental setup to be unclear.

Critical weaknesses
  1. The authors' motivation is unclear; why do we need a DNBBA mechanism?
  2. The description of the DNBBA algorithm itself is quite abstract, making it unclear how it is applied to specific tasks; the authors merely provide a loss function and the BBA processing method, without even explaining the origin of BBA.
  3. Furthermore, the experiments appear disjointed, involving two scenarios that lack a strong connection. Are the authors attempting to demonstrate the method's applicability across a wide range of scenarios?
Questions for authors
  1. The authors' motivation is unclear; why do we need a DNBBA mechanism?
  2. The description of the DNBBA algorithm itself is quite abstract, making it unclear how it is applied to specific tasks; the authors merely provide a loss function and the BBA processing method, without even explaining the origin of BBA.
  3. Furthermore, the experiments appear disjointed, involving two scenarios that lack a strong connection. Are the authors attempting to demonstrate the method's applicability across a wide range of scenarios?
Reviewer_txknOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes dynamic negation of basic belief assignments, which redistributes belief using kernel determined by proposition differences and current belief values.

Negated BBAs are combined using Dempster Shafer's rule with classification based on minimum negated belief. The framework is also used to guide low-light image enhancement. The paper presents theoretical properties and reports improvements across classification and enhancement benchmarks.

Critical strengths

Exploring adaptive belief redistribution within learned decision systems is potentially interesting. The evaluation covers multiple appliocations, and Tables 5-6 provide initial component ablations. The reported classification gains and improvements in enhancement SSIM motivates further investigations.

Critical weaknesses

My main concerns concern the validity of the theoretical claims and whether the experiments establish the contribution of adaptive negation.

  1. Several general theoretical claims are incorrect or inadequately justified. Property 3 treats the transformation kernel as fixed in Eq. (27), although the definition makes it depend on the input beliefs: T(m)=Ψ(m)mT(m)=\Psi(m)m. The proof therefore does not establish the claimed linearity. Property 6’s entropy argument also uses Jensen’s inequality incorrectly; column normalization alone does not guarantee increasing entropy.
  2. The convergence analysis does not establish the advertised general behavior. Theorem 3 analyzes the singleton recurrence T(m)i=(1−mi)/(C−1)T(m)_i=(1-m_i)/(C-1), rather than the general adaptive operator. Its convergence to uniformity requires C>2C>2; binary distributions generally oscillate.
  3. The experimental contribution of the adaptive kernel is unclear. If the model operates only on singleton propositions, the two set-difference terms equal one for distinct classes, so α+β=1\alpha+\beta=1 and ρ\rho cancels completely. The paper needs to specify the proposition families used in every experiment. Table 5 does not distinguish ordinary negation from adaptive negation or changes in fusion and training. Table 6 is more relevant, but compares learned ρ\rho against only one explicitly specified fixed value and reports no variability across runs. Missing controls include ordinary complement negation, existing BBA-negation operators, fixed kernels, and simple learned reweighting under identical training conditions.
  4. The experiments do not directly test the principal claimed benefits. Standard accuracy does not establish improved uncertainty estimation, conflict handling, or resistance to class imbalance. Controlled experiments varying evidence conflict, imbalance, and label ambiguity are needed, together with appropriate reliability metrics. For enhancement, the results show a PSNR–SSIM tradeoff: PSNR is below Retinexformer on LOL-v1 and LOL-v2-real, while model complexity increases. An enhancement-specific comparison against generic spatial weighting or attention is missing.
  5. Important implementation details and relevant comparisons are absent. The paper does not adequately specify BBA construction, focal-set selection, the classification training objective, constraints on ρ\rho, or the enhancement degradation map. Equation (11)’s empty-set assignment contradicts the BBA definition, and zero normalization denominators require explicit handling. Existing negation-based methods need substantive comparison. Fan et al. (2025), already cited and addressing closely related classification and enhancement tasks, is particularly relevant but absent from the quantitative comparisons.
Questions for authors
  1. Which propositions and evidence sources are instantiated for each task? When and how do nonsingleton propositions arise?
  2. What differentiable objective trains ρ\rho, how is it constrained, and how are empty-set/full-frame masses and zero denominators handled?
  3. Can the authors provide matched comparisons against ordinary negation, prior BBA-negation methods, a tuned fixed kernel, and simple learned weighting, with multiple seeds?
  4. Can targeted conflict/imbalance experiments and enhancement-specific ablations establish the claimed benefits beyond overall benchmark accuracy?
Reviewer_1QntOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes DNBBA, a mass-dependent asymmetric operation for redistributing basic belief assignments. The negated assignments are combined using Dempster's rule, and classification is performed by selecting the class with the smallest fused negative mass. The method is also applied to low-light image enhancement, where the resulting assignments are used to weight image inputs. The paper presents several mathematical properties and evaluates the method on classification and image enhancement tasks, together with component ablations.

Critical strengths
  • Eqs. (8)–(10) provide an explicit formulation of negative-belief revision through a set-difference-based redistribution. This makes the proposed operation concrete and allows its properties to be examined mathematically.

  • The ablation experiments provide useful evidence for the contributions of the proposed components. In Table 5, CIFAR-10 accuracy increases from 93.25% for the baseline to 94.97% with BBA and 96.12% with DNBBA. Table 6 further compares fixed and learned ρ\rho settings under partial-label learning. These comparisons help isolate the effects of the proposed components.

Critical weaknesses
  • Several important mathematical claims appear to be incorrect or require additional assumptions. Eq. (11) assigns unit mass to the empty set, contradicting the BBA definition in Eq. (4), and the same issue appears in Appendix A.2. Corollary 1 also omits the condition ∣Θ∣>2|\Theta|>2: for two singleton classes, the stated recursion alternates between (0.8,0.2)(0.8,0.2) and (0.2,0.8)(0.2,0.8) instead of converging to a uniform distribution. In addition, the proof of the affine property in Eq. (27) treats the redistribution kernel as fixed, even though Eq. (11) makes it dependent on the input masses. These issues affect the theoretical justification of DNBBA and cannot all be resolved as typographical errors.

  • The description of DNBBA-DNet lacks several implementation details needed to understand and reproduce the reported results. In particular, the paper does not clearly define how features are mapped to focal-set masses, how the learnable ρ\rho is constrained, or which loss is used to train the classifier. For low-light image enhancement, Eqs. (14)–(15) also leave the degradation map and enhancement objective insufficiently specified. Although Appendix A.3 provides dataset information and hardware details, it does not resolve these questions. As a result, it is difficult to determine how the proposed operation produces the reported improvements.

Questions for authors
  • Is the empty-set assignment in Eqs. (11) and (21) a typographical error, and which focal sets are included in the normalization? Could the authors also clarify the assumptions required for the convergence and affine properties, particularly for ∣Θ∣=2|\Theta|=2 and when the redistribution kernel depends on mm? Corrected statements and proofs would help resolve the theoretical concerns.

  • Could the authors provide more details on how features are converted into BBAs, which focal sets are used, how ρ\rho is constrained, and how the classification and enhancement losses are defined? Please also clarify the construction of partial-label candidates and the training budgets used for Tables 5–6. These details are necessary to assess the implementation and experimental comparisons.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.