acceptodds
Under review as a conference paper at ICLR 2027

Fisher-Consistent Evidential Learning under Prior Shift

Abstract

Uncertainty representations are useful only when the decisions they induce remain coherent as prior beliefs change. Under long-tailed training, the mean–variance core of standard evidential deep learning (EDL) preserves the ordering of the training-prior posterior; under an explicit ordering-stability condition, its population-optimal classifier therefore inherits the training log-prior term. In a shared-covariance Gaussian model, this term yields a closed-form boundary displacement and can eliminate every neighborhood of a minority-class mean. We propose Fisher-Consistent Evidential Learning (FC-EDL), which transports a strictly proper mean risk to a chosen reference prior by importance weighting and controls Dirichlet concentration through a separate regularizer. The resulting mean branch is Fisher-consistent for the reference-prior posterior, while the concentration branch retains a valid Dirichlet uncertainty representation. Across diverse classification settings spanning vision, video, text, and fixed-label LLMs, FC-EDL consistently improves minority-class performance and calibration, substantially mitigates the adverse effects of explicit prior shift, and yields robust, more decisive uncertainty representations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Evidential Deep Learning, Uncertainty Quantification, Long-Tailed Recognition, Prior Shift, Fisher Consistency

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_SmfYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The main idea of the paper is to investigate the influence of imbalanced training priors on classification decisions. The theoretical formulation part is professional, and the results are promising. However, there are some issues related to the methodology, novelty, experiments, and uncertainty modeling.

Critical strengths
  1. The paper addresses the impact of class imbalance on the minority-class decision boundaries. 2.Its analysis includes fisher consistency, importance-weighted risk minimization, the gaussian boundary displacement, and minority-region problem.
  2. The predictive separation from dirichlet makes a simple framework to train the uncertainty-aware classifiers.
  3. The evaluation includes different modalities and fixed-label LLM classification
  4. The evaluation includes several few-shot groups in which the majority-class performance is maintained as well.
Critical weaknesses

1- It is very importatn to distigush better the contribution and main novelty of the paper from the past and conventional class re-weithing methods.

  1. There are some assumptions that limit the generality of the model. for example the theorem 3 assumes known priors, pure prior shift, and independent optimization. Also the theorem 1 requires an ordering-stability condition .

  2. There are some mismatchs between theory and the implementation. The theorem 3 shows exact importance weighting, whereas the implementation uses clipped weights and minibatch self-normalization. Fixed clipping can change the effective target then the theoretical guarantee does not directly apply.

  3. There are also some insufficient experimental controls. To determine whether improvement comes from conventional reweighting rather than from the evidential formulation, a class-weighted cross-entropy model with the same head is needed. You should also add more complete baseline variability and statistical significance.

  4. Examples include “Equation equation 13”, “The paper provide”, and repeated equation-reference wording in the supplementary section.

  5. The Cao et al. (2019) reference appears twice with identical bibliographic information. Also, there are several references contain incomplete author formatting or extraneous text.

  6. Tables 1 and 2 are dense and some figure labels are small.

  7. Important selected hyperparameter values should also be provided.

Questions for authors

1- please clear the main novelty of your work versus the other classical methods.

2- Please clear the generality of your work, in the appreance of many limitations and constraints you defined.

2 comments

Regarding weight clipping, please refer to lines 266–269 of the manuscript. It is solely intended to prevent training explosion caused by extremely large weights during training; in experiments, a factor of 10 is already a quite exaggerated threshold.

Assuming known priors only confirms the correctness of the prior information (all of which comes from the target domain), and has nothing to do with generalizability. The theme of this paper is evidential optimization of Fisher consistency theory on long-tailed datasets. The same head classes and baselines have also been given in the paper (see Table 1 and Table 2), and the comparison is between the accuracy of the original baseline and that after Fisher optimization.

Reviewer_9TZ4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the effects of class-prior shift on Evidential Deep Learning (EDL) and proposes Fisher-Consistent Evidential Learning (FC-EDL). The method employs importance weighting to align the predictive-mean objective with a reference prior while separately regularizing Dirichlet concentration. The authors provide theoretical analyses of prior-induced decision-boundary shifts and reference-prior Fisher consistency, along with experiments across vision, video, text, and LLM classification tasks, demonstrating improvements in long-tailed recognition and predictive calibration.

Critical strengths
  1. The paper investigates the relationship between class-prior shift, classification decisions, and evidential uncertainty representations, addressing an important issue in long-tailed learning.
  2. The paper provides a mathematical characterization of how class priors influence decision boundaries, including a quantitative condition for local minority-region collapse under Gaussian assumptions.
  3. The experiments cover diverse datasets, architectures, and modalities. The reported improvements in minority-class recognition and predictive calibration are encouraging.
Critical weaknesses
  1. The paper introduces several abstract terms, such as "decision-inconsistent uncertainty representation," "population-optimal decision rule," and "boundary collapse," without sufficiently explaining their meanings. More importantly, the relationship between prior-induced classification bias and unreliable evidential uncertainty estimation is not clearly established. The authors should better distinguish these two problems and clarify their connection.
  2. The concentration regularizer, (log⁡S−τ)2(\log S-\tau)^2, encourages the concentration to approach the same target value for all inputs, regardless of their prediction reliability or uncertainty. This raises concerns about whether the concentration branch can learn meaningful input-dependent uncertainty.
  3. Although FC-EDL separates predictive-mean and concentration estimation, its advantages over the standard EDL parameterization are not sufficiently demonstrated. Since the proposed method also modifies the training objectives, it remains unclear whether the improvements arise from parameter decoupling or importance-weighted classification.
  4. Although the paper reports extensive improvements in classification accuracy and calibration, these results do not directly establish whether the new training objective and parameterization improve evidential uncertainty estimation. Table 3 provides useful ablations, but the effects on Dirichlet concentration and uncertainty quality remain unclear. Additional evaluations of input-dependent uncertainty, error detection, and OOD detection would better substantiate the claimed benefits of evidential learning.

Minor Weakness:

  1. Despite the extensive theoretical motivation, FC-EDL primarily combines importance-weighted cross-entropy with a separately regularized Dirichlet concentration head. Since importance weighting is a well-established technique for prior correction, the authors should clarify the methodological novelty beyond conventional importance-weighted classification and explain the additional benefits introduced by the evidential formulation.
  2. In Line 254, "Equation equation 13" contains a duplicated word and should be corrected to "Equation (13)."
Questions for authors

Please refer to the Weakness part.

1 comment

Regarding the definitions of Fisher Consistency and Prior-induced Boundary Collapse, please refer to lines 113–161 of the manuscript. The paper has already provided the mathematical definition of Fisher Consistency, the prior-coupling condition of Standard EDL, and the theoretical derivation of the collapse of minority-class decision regions. Regarding the modeling scope of the concentration branch, please refer to lines 262–309 of the manuscript. This paper achieves the decoupling of the predictive mean and the total concentration through independent parameterization, but does not interpret the current concentration regularization as an independent estimator of input-dependent epistemic uncertainty. Regarding the effectiveness of parameter decoupling, please refer to Equations 10 and 11, lines 289–299, and Table 3 of the manuscript. Experiments show that compared with Coupled α\alpha, independent mean-concentration parameterization achieves better results in overall accuracy, minority-class accuracy, and calibration metrics.

Reviewer_wDADOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how training priors affect the decision boundaries of evidential classifiers and proposes FC-EDL to address this issue. The method applies reference-prior weighting to a proper loss for the predictive mean while regularizing Dirichlet concentration separately. The authors establish Fisher consistency under label shift and a realizable function class in which the predictive mean and concentration can vary independently. They also provide Gaussian boundary analysis and experiments on long-tailed classification and prior shift to demonstrate the effectiveness of the proposed method.

Critical strengths
  • The paper clearly defines the target prior for decision-making. Eq. (6) shows how importance weighting transforms the mean risk to the reference-prior distribution, while Theorem 3 establishes Fisher consistency under the stated assumptions. This provides a clear theoretical justification for the proposed correction.

  • The experiments cover a broad range of classification tasks, with particularly strong improvements for minority classes. For example, on ImageNet-LT with ResNet-50, Few accuracy improves from 34.0% to 46.8%, while overall accuracy increases from 51.9% to 53.8%. The ablation studies and prior-shift experiments provide further evidence of the method's effectiveness.

  • The Gaussian analysis provides an intuitive explanation of how class imbalance can affect decision boundaries. In particular, Corollary 2 derives the condition log⁡(πj/πk)≥dkj2/(2σ2)\log(\pi_j/\pi_k) \geq d_{kj}^2/(2\sigma^2), under which the minority-class mean loses its surrounding decision region. This helps connect the theoretical analysis to the motivation for correcting training-prior bias.

Critical weaknesses
  • The concentration regularizer does not explain how the model learns input-dependent evidential uncertainty. Under the population-level assumptions, minimizing (log⁡S−τ)2(\log S-\tau)^2 with a positive regularization weight gives the same optimal concentration S=exp⁡(τ)S=\exp(\tau) for every input. Although the resulting Dirichlet distribution is valid, it is unclear whether the concentration branch captures meaningful uncertainty beyond the predictive mean. The authors should provide additional evidence of input-dependent concentration behavior and clarify whether the calibration improvements in Table 3 come from this branch or from changes in the predictive mean and shared representation.

  • The decision boundaries shown in Figure 3 are estimated using Gaussian prototypes in a projected feature space, rather than directly from the trained classifier. Therefore, the visualization does not necessarily demonstrate that the model's actual decision boundaries have been recovered. In addition, the discussion on p. 9 describes Figure 3 as a sweep over (δ,log⁡ρ)(\delta,\log\rho), which does not match the displayed figure. The authors should clarify the visualization procedure and provide more direct evidence for the claimed boundary recovery.

  • The reported classification improvements may be largely explained by reference-prior importance weighting, which is already well established in learning under distribution shift. Although FC-EDL combines this approach with a decoupled Dirichlet representation, it is unclear how much the evidential formulation contributes beyond a conventional classifier trained with the same weighted loss. A comparison with a standard cross-entropy classifier using identical reference-prior weights would help isolate the benefit of the proposed formulation.

Questions for authors
  • Given that the concentration penalty in Section 4 has a constant optimum at S=exp⁡(τ)S=\exp(\tau), what evidence shows that S(x)S(x) captures meaningful input-dependent uncertainty? Please clarify how λ\lambda and τ\tau are selected and whether the reported calibration improvements can be attributed specifically to the concentration branch rather than changes in the predictive mean or shared backbone.

  • The discussion on p. 9 refers to a parameter sweep over (δ,log⁡ρ)(\delta,\log\rho) that does not appear to match Figure 3. Could the authors clarify where this analysis is presented? Also, are the actual network decision boundaries evaluated independently of the Gaussian prototype approximation, and how is the target prior incorporated into the post-hoc correction baseline in Table 4?

2 comments

Regarding the visualization in Fig. 3, please refer to Lines 425–440 of the manuscript. During actual training and validation, a monitor callback records each sample's latent representation, predictive mean, and assigned label at every checkpoint. The visualization is constructed based on these checkpoint-level records.

Regarding the scope of the concentration regularization term, please refer to lines 233–237 of the manuscript. The paper has explicitly distinguished between the predictive mean and the total concentration, and stated that the current regularization term is used to control the concentration scale, rather than independently identifying input-dependent epistemic uncertainty. Regarding hyperparameter selection, please refer to lines 321–332 of the manuscript. All hyperparameters are selected on the validation set and fixed before testing, without using test set metrics for hyperparameter selection. Regarding the role of concentration regularization, please refer to Table 3 and lines 463–474 of the manuscript. After removing concentration regularization, ECE increases from 0.035 to 0.082, and SelectiveRisk increases from 0.120 to 0.205, indicating that this component makes an additional contribution to model calibration and selective prediction. The conventional classifiers trained with the same weighted loss are directly derived from the baselines.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.