acceptodds
Under review as a conference paper at ICLR 2027

Embracing Aleatoric and Epistemic Uncertainty for Trusted Open-world Multi-view Classification

Abstract

Multi-view classification fuses heterogeneous information to achieve more comprehensive data descriptions. To deliver trustworthy decisions, trusted multi-view classification (TMVC) has emerged as a crucial research paradigm. However, existing TMVC methods assume closed-set environments and inevitably fail when novel categories emerge in open-world settings. This stems from two key limitations: they neglect data-inherent aleatoric uncertainty, leaving cross-view alignment vulnerable to noise; and their fusion strategies lack epistemic coherence, producing overconfident and unreliable predictions that are catastrophically amplified for unknown samples. To address these, we propose a novel framework jointly Embracing aleatoric and epistemic Uncertainty for trusted Open-world multi-view classification (EUO). Specifically, we quantify view-specific aleatoric uncertainty and align cross-view representations to obtain a robust joint representation, which serves as a reliable reference for subsequent opinion aggregation. Besides, we propose an Uncertainty-Driven Proportional Conflict Redistribution (UD-PCR) operator that aggregates multi-view subjective opinions by adaptively redistributing conflict mass in proportion to calibrates belief masses and uncertainty. We also theoretically prove its effectiveness, particularly in open-world scenarios. A novel Jousselme consistency loss further enforce cross-view epistemic alignment at a linear complexity. Experiments on multiple benchmarks demonstrate superior closed-set accuracy and open-world rejection.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Trusted multi-view classification, Open-set recognition, Multi-view fusion, Uncertainty estimation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_gb1GOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on open-world multi-view classification, simultaneously modeling the aleatoric uncertainty induced by noisy data and the epistemic uncertainty arising from unknown classes. The method represents each view with a Gaussian distribution, leverages KAN to collect evidence, and proposes UD-PCR to resolve cross-view conflicts. Closed-set and open-set experiments on six datasets demonstrate competitive overall performance.

Critical strengths
  1. The aleatoric uncertainty at the representation level and the epistemic uncertainty at the decision level are addressed within a unified framework.
  2. UD-PCR comes with a clear mathematical formulation and an analysis of its boundary properties. Experiments span six datasets and report OSCR, Open AUC, and low-FPR metrics, yielding a relatively comprehensive evaluation.
Critical weaknesses
  1. The open-set experiments are compared mainly against closed-set trustworthy multi-view methods, whereas open-world multi-view approaches discussed in the related work, such as OpenViewer and TOMC, are not included. Could the authors supplement this section with additional comparisons?
  2. The variance produced by the Gaussian encoder is directly interpreted as aleatoric uncertainty, yet no calibration metrics, correlation analysis with noise intensity, or quantitative perturbation experiments are provided to substantiate this interpretation.
Questions for authors
  1. How are known and unknown classes partitioned? Do the five runs adopt the same class split, and is an independent validation set used for model selection and threshold determination?
  2. Could the authors quantitatively verify the correlation between the Gaussian variance and input noise or prediction errors, and supplement an ablation study in which KAN is replaced solely by an MLP?
3 comments
AuthorsOpenReview ↗

Author Response to Reviewer gb1G — Part 1

We sincerely thank you for your careful evaluation and constructive suggestions, which have helped us improve the clarity and completeness of our manuscript. We have carefully addressed your concerns through additional experiments, methodological clarifications, and corresponding revisions. Our detailed responses are provided below.

[W1] The open-set experiments are compared mainly against closed-set trustworthy multi-view methods, whereas open-world multi-view approaches discussed in the related work, such as OpenViewer and TOMC, are not included. Could the authors supplement this section with additional comparisons?

We sincerely thank you for this valuable suggestion.

(1) Comparison with OpenViewer.

Our initial comparisons primarily focus on existing TMVC methods, as EU²O aims to achieve reliable uncertainty estimation and trustworthy decision-making. OpenViewer [1] primarily addresses unknown-category recognition by generating pseudo-unknown samples to mitigate overconfident predictions.

To provide a more comprehensive evaluation, we further include OpenViewer as an additional open-set baseline. For fairness, we adapt OpenViewer to our unified experimental protocol (e.g., identical data partitions), while preserving its original architecture, loss functions, and core learning mechanisms.

Table 1. Comparison with OpenViewer under the unified experimental protocol.

DatasetMethodOSCR ↑Open AUC ↑
Flower17OpenViewer17.33±0.6051.13±7.34
EU²O76.79±8.1879.73±5.76
VGGFace2-50OpenViewer61.33±12.5175.80±7.47
EU²O71.80±2.1979.57±1.85
YouTubeOpenViewer36.21±14.6256.79±9.12
EU²O78.93±4.6181.43±3.98
SceneOpenViewer21.81±5.7048.99±7.66
EU²O68.13±12.7871.19±10.74
MVSA-SingleOpenViewer36.09±15.8454.59±6.50
EU²O56.22±3.2463.16±1.63
NYU Depth V2OpenViewer40.51±10.3847.30±10.46
EU²O48.96±10.0153.36±9.80

As shown in Table 1, EU²O achieves higher mean OSCR and Open AUC across all six benchmarks under the unified protocol, further demonstrating its effectiveness in trusted open-set recognition.

(2) Discussion of TOMC [2].

Its official implementation was unavailable to us, and certain implementation details could not be reliably recovered from the paper alone. We discuss TOMC and its methodological differences in the Related Work section and further clarify these distinctions in the revision.

References:

[1] OpenViewer: Openness-Aware Multi-View Learning, ACM MM 2025.

[2] Trusted Open-World Multi-View Classification with Dynamic Opinion Aggregation, ACM MM 2025.

AuthorsOpenReview ↗

Author Response to Reviewer gb1G — Part 2

[W2] The variance produced by the Gaussian encoder is directly interpreted as aleatoric uncertainty, yet no calibration metrics, correlation analysis with noise intensity, or quantitative perturbation experiments are provided to substantiate this interpretation.

Thank you for this valuable suggestion. We provide further clarification and additional experiments to substantiate the learned Gaussian variance as an indicator of view-specific aleatoric uncertainty.

(1) Aleatoric uncertainty modeling and noise correlation.

Our approach follows established probabilistic representation learning methods (Shi and Jain, ICCV 2019 [1]; Tang et al., ACL 2025 [2]) and recent distribution-based aleatoric uncertainty modeling (Tang et al., IJCAI 2025 [3]; Wang et al., CVPR 2026 [4]). Specifically, in our framework, each sample is mapped by encoders into a mean vector and a standard deviation vector, which jointly parameterize its latent Gaussian distribution. The mean captures view-specific semantic information, while the variance characterizes uncertainty associated with noisy or ambiguous observations, providing a principled basis for modeling view-specific aleatoric uncertainty.

Fig. 1. Analysis of Gaussian Variance and Adaptive Cross-View Fusion under Increasing Input Noise.

To examine its relationship with input noise, we introduced varying levels of Gaussian noise into one view of the Scene dataset. As shown in Fig. 1(a–b), the change in predicted variance exhibits a positive correlation with noise intensity (Pearson r=0.214r=0.214), while the average variance of the corrupted view increases with noise severity. Fig. 1(c–d) further shows that AU-DCF adaptively reduces the corrupted view's contribution and achieves better OSCR and Open AUC than uniform fusion.

(2) Quantitative perturbation experiments.

We further conducted experiments under varying Gaussian noise intensities (0–1.5), comparing EU²O with six representative TMVC baselines.

Fig. 2. Robustness Comparison under Varying Gaussian Noise Intensities on Three Datasets.

As shown in Fig. 2(a–b), EU²O generally demonstrates favorable robustness in both closed-set classification (Accuracy) and open-set recognition (OSCR and Open AUC) as noise intensity increases.

In summary, these results provide empirical support for the noise sensitivity of the learned Gaussian variance and the robustness of AU-DCF. We have incorporated these analyses into the revised manuscript.

References:

[1] Probabilistic Face Embeddings, ICCV 2019.

[2] Modeling Uncertainty in Composed Image Retrieval via Probabilistic Embeddings, ACL 2025.

[3] DFMU: Distribution-based Framework for Modeling Aleatoric Uncertainty in Multimodal Sentiment Analysis, IJCAI 2025.

[4] Uncertainty-Aware Modality Fusion for Unaligned RGB-T Salient Object Detection, CVPR 2026.

AuthorsOpenReview ↗

Author Response to Reviewer gb1G — Part 3

[Q1] How are known and unknown classes partitioned? Do the five runs adopt the same class split, and is an independent validation set used for model selection and threshold determination?

Thank you for the question. We clarify our data partitioning and evaluation protocol as follows.

(1) Known/unknown class partitioning.

Following the standard openness definition

O=1−2KK+N,O = 1 - \sqrt{\frac{2K}{K+N}},

we determine the number of known classes KK from the total number of classes NN and the target openness. The known classes are randomly selected, with 80% of their samples used for training and the remaining 20% for testing. All samples from unknown classes are exclusively reserved for testing and are never used during training.

(2) Five independent runs.

We perform five independent random partitions rather than using a fixed class split across all runs. Importantly, within each run, all competing methods share the same known/unknown class partition and training/test split to ensure fair comparisons. We report the mean and standard deviation across the five runs.

(3) Model selection and threshold determination.

We do not use an independent validation set for checkpoint selection or rejection-threshold tuning. Instead, each model is evaluated after completing its predefined training schedule. OSCR and Open AUC are computed by sweeping confidence thresholds on the held-out test predictions, while CCR@FPR is reported at predefined FPR levels (0.5% and 1%), rather than using a separately tuned deployment threshold.

We have clarified these implementation details in the revised manuscript.

[Q2] Could the authors quantitatively verify the correlation between the Gaussian variance and input noise or prediction errors, and supplement an ablation study in which KAN is replaced solely by an MLP?

Thank you for the suggestion. We have conducted additional experiments to address both concerns.

(1) Correlation between Gaussian variance and input noise.

We introduced Gaussian noise of varying intensities into View 3 of the Scene dataset. As shown in Fig. 1(a), the change in predicted variance exhibits a modest positive correlation with noise intensity (Pearson r=0.214r=0.214). Moreover, the average predicted variance increases with noise severity (Fig. 1(b)), while AU-DCF adaptively reduces the corrupted view's fusion weight (Fig. 1(c)). Fig. 1(d) further demonstrates better Open AUC than uniform fusion under increasing noise.

These results provide empirical support for the noise sensitivity of the learned Gaussian variance, enabling AU-DCF to adaptively mitigate the influence of unreliable views and improve the robustness of cross-view fusion.

(2) KAN-to-MLP ablation study.

We further replaced only the KAN component with an MLP while keeping all other modules and experimental settings unchanged. The results on Scene are summarized below.

Table 2. Ablation study of the KAN-based evidential collector on Scene under different openness settings.

OpennessCollector[email protected]% ↑CCR@1% ↑OSCR ↑Open AUC ↑
0.2MLP5.28±4.327.70±6.2161.73±4.9966.62±5.67
KAN5.54±4.038.42±5.7362.08±4.2467.62±5.26
0.3MLP7.70±4.099.78±4.8167.47±12.3070.33±10.56
KAN7.82±4.9710.64±7.1767.62±13.0570.54±11.09

As shown in Table 2, KAN achieves higher performance across the evaluated metrics on Scene under different openness settings, providing additional improvements over the MLP-based collector.

We have incorporated these additional analyses and ablation results into the revised manuscript.


We hope our rebuttal clarifies and addresses all your concerns, and we also welcome any further constructive feedback you might provide to enhance our submission.

Reviewer_8zpSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the EU²O framework for open-world trustworthy multi-view classification. The method first represents features from different views as Gaussian distributions incorporating uncertainty and fuses them based on view uncertainty; subsequently, it employs an evidential network to transform view-specific features into class-level evidence and models prediction uncertainty using a Dirichlet distribution. Finally, it introduces UD-PCR for multi-view evidence fusion, adjusting the allocation of conflicting information based on view uncertainty while enforcing consistency constraints to align predictions across views, thereby enabling the classification of known categories and the rejection of unknown ones.

Critical strengths

This paper investigates the issue of uncertainty modeling in open-world multi-view classification and validates the approach through experiments on multiple multi-view and multi-modal benchmarks.

Critical weaknesses
  1. The paper claims to study open-world multi-view classification. However, the task is mainly open-set recognition, where the model only needs to detect unknown classes. It does not consider learning new classes over time. Thus, the title and motivation are broader than the actual problem and experiments.

  2. The contribution of the proposed fusion method is unclear. The performance gains may come from KAN, the fusion method, or other parts of the framework. Although Table 5 compares different fusion methods using the same MLP, the gains are small and are only tested on one dataset. More controlled experiments are needed to show the separate effects of KAN and the proposed fusion method.

  3. The novelty of the method is limited. Many existing studies have already combined aleatoric and epistemic uncertainty for multi-view learning. Evidence-based fusion and conflict handling have also been widely studied. The paper does not clearly explain how the proposed method differs from these methods or why it is better.

  4. The proposed consistency loss is not fully studied. The authors introduce a Jousselme consistency loss to improve agreement across views. However, the Jousselme distance is an existing measure, and its advantage over common losses such as MSE, KL divergence, or JS entropy is not shown. Table 3 only provides a limited ablation study. Also, the notation (L_{Jou}) in Table 3 is not consistent with (L_{con}) in the main text.

  5. Most theoretical results only consider extreme cases where uncertainty approaches 0 or 1. The paper does not show whether these cases are close to the outputs of the trained model. It also lacks analysis for finite evidence values. More importantly, Proposition 3 only considers two-view fusion, while the proposed method uses sequential multi-view fusion. Adding another confident opinion may greatly reduce the high uncertainty caused by two conflicting views. Thus, the current theory cannot guarantee that the final model preserves high uncertainty under view conflicts or for unknown samples. Also, Eq. (12) is undefined when both input opinions have uncertainty 1, which is not covered in Proposition 1. These issues weaken the claimed theoretical support for the proposed fusion method.

  6. Only one of the six datasets has more than 10,000 samples. Although the appendix provides some examples with noise and view conflicts, these are not enough to show the robustness of the method. More tests with different noise levels and conflict levels are needed. Also, uncertainty calibration should be evaluated using metrics such as ECE, NLL, or Brier Score. Comparisons with direct open-set multi-view methods, such as TOMC and OpenViewer, are also missing.

Questions for authors

See weaknesses.

7 comments
AuthorsOpenReview ↗

Author Response to Reviewer 8zpS — Part 1

We sincerely thank you for your careful evaluation and constructive suggestions, which have helped us improve the clarity and completeness of our manuscript. We have carefully addressed your concerns through methodological clarifications, additional experiments, theoretical analysis, and corresponding revisions. Our detailed responses are provided below.

[W1] The paper claims to study open-world multi-view classification. However, the task is mainly open-set recognition, where the model only needs to detect unknown classes. It does not consider learning new classes over time. Thus, the title and motivation are broader than the actual problem and experiments.

Thank you for highlighting this important distinction. We would like to clarify that EU²O focuses on reliable known-class classification and unknown-class rejection in an open-world multi-view environment, rather than incremental learning of newly discovered categories.

While the broader formulation of open-world recognition also encompasses incremental class learning [1], unknown-category recognition constitutes an important component of this setting. Recent work, such as Trusted Open-World Multi-View Classification with Dynamic Opinion Aggregation [2], similarly investigates unknown-category detection in open-world multi-view scenarios.

To clarify the scope of our contribution, we distinguish the open-world learning environment from the open-set recognition task evaluated in this work and refine the corresponding motivation and problem formulation in the revised manuscript.

References:

[1] Towards Open World Recognition, CVPR 2015.

[2] Trusted Open-World Multi-View Classification with Dynamic Opinion Aggregation, ACM MM 2025.

AuthorsOpenReview ↗

Author Response to Reviewer 8zpS — Part 2

[W2] The individual contributions of KAN and the proposed fusion method remain unclear. More controlled experiments across datasets are needed to isolate their effects.

Thank you for this constructive suggestion. We further clarify the individual contributions of UD-PCR and KAN through two controlled comparisons.

(1) Effectiveness of UD-PCR.

As described in Appendix A.5.2, we compare different fusion strategies under a unified MLP-based framework, excluding the influence of KAN, AU-DCF, and the consistency loss. To further assess generalizability, we extend these experiments to Scene and YouTube under openness = {0.3, 0.4}, while keeping the remaining experimental settings unchanged.

Table 1. Comparison of different fusion strategies on Scene and YouTube under different openness settings. Lower Avg. Rank indicates better performance.

DatasetOpennessFusion[email protected]% ↑CCR@1% ↑OSCR ↑Open AUC ↑Avg. Rank ↓
Scene0.3MEAN5.01±4.288.81±6.0461.12±14.0265.11±12.493.50
TMC4.84±3.808.17±4.9661.57±13.7065.43±12.113.75
RCML5.13±3.708.52±5.8361.01±13.1565.49±11.142.75
TMCEK5.03±3.669.08±6.5560.99±13.4065.35±11.473.25
UD-PCR5.38±3.848.45±5.0461.96±13.8365.81±12.201.75
0.4MEAN7.51±7.0910.55±8.6959.88±8.3962.50±8.204.75
TMC7.56±6.5310.66±8.7960.64±8.1863.14±8.112.50
RCML8.17±6.5911.37±8.9160.34±8.2863.11±7.882.00
TMCEK7.81±6.3310.75±8.3260.32±8.2163.08±7.793.00
UD-PCR7.48±6.4510.64±8.5061.46±7.9464.04±7.682.75
YouTube0.3MEAN8.83±5.3914.83±8.9068.36±8.5474.50±6.412.25
TMC8.67±5.0215.33±8.3167.11±10.0873.70±7.352.75
RCML6.33±7.0210.50±10.8262.09±10.9968.25±8.864.00
TMCEK5.67±6.6510.00±10.2261.96±10.7968.16±8.705.00
UD-PCR9.67±4.5516.33±11.0468.75±8.5674.65±6.601.00
0.4MEAN2.25±1.663.75±3.2659.65±9.4662.33±10.291.50
TMC1.75±1.003.25±2.3258.21±9.2861.27±9.982.88
RCML0.50±0.610.75±0.6151.27±10.3154.19±11.764.88
TMCEK0.50±0.611.00±0.9451.52±9.9654.47±11.464.13
UD-PCR2.00±1.273.25±2.8160.75±8.8163.45±9.501.63

As shown in Table 1, UD-PCR achieves the highest OSCR and Open AUC across all four settings, demonstrating its effectiveness in open-set recognition.

Table 2. Average ranking of different fusion strategies across datasets and openness levels (lower is better).

DatasetOpennessMEANTMCRCMLTMCEKUD-PCR
Scene0.33.503.752.753.251.75
0.44.752.502.003.002.75
YouTube0.32.252.754.005.001.00
0.41.502.884.884.131.63
Overall Avg. Rank ↓—3.002.973.413.841.78
Final Rank ↓—32451

Notably, UD-PCR achieves the best overall average rank of 1.78 (final rank: 1), demonstrating that its effectiveness does not depend on the KAN-based evidence collector.

(2) Additional Contribution of KAN.

To isolate the effect of KAN, we replace the KAN-based evidence collector with an MLP while keeping UD-PCR and the remaining model components unchanged. The results on Scene are summarized in Table 3.

Table 3. Ablation study of the KAN-based evidential collector on Scene under different openness settings.

OpennessCollector[email protected]% ↑CCR@1% ↑OSCR ↑Open AUC ↑
0.2MLP5.28±4.327.70±6.2161.73±4.9966.62±5.67
KAN5.54±4.038.42±5.7362.08±4.2467.62±5.26
0.3MLP7.70±4.099.78±4.8167.47±12.3070.33±10.56
KAN7.82±4.9710.64±7.1767.62±13.0570.54±11.09

As shown in Table 3, KAN achieves higher mean performance across all evaluated metrics on Scene under different openness settings, providing additional improvements over the MLP-based collector.

In summary, these controlled experiments distinguish the contributions of the proposed components: UD-PCR provides an effective opinion fusion mechanism in open-set recognition, while KAN offers complementary improvements in evidence-based classification. We have included these additional comparisons and discussions in the revised manuscript.

AuthorsOpenReview ↗

Author Response to Reviewer 8zpS — Part 3

[W3] The novelty of the method is limited. Many existing studies have already combined aleatoric and epistemic uncertainty for multi-view learning. Evidence-based fusion and conflict handling have also been widely studied. The paper does not clearly explain how the proposed method differs from these methods or why it is better.

Thank you for this insightful comment. We would like to clarify that the contribution of EU²O lies in addressing two fundamental limitations of existing trusted multi-view classification methods in open-world scenarios.

(1) Limitations in Representation-Level Uncertainty Modeling.

Representative TMVC methods, such as TMC, CCML, RCML and GTMC, primarily focus on decision-level evidence aggregation, overlooking the impact of view-specific aleatoric uncertainty on representation learning. EU²O explicitly models aleatoric uncertainty to achieve robust cross-view alignment and fusion, further constructing a reliable joint opinion to guide the final opinion aggregation.

(2) Limitations of Conventional Opinion Fusion in Open-Set Recognition.

As theoretically established in Appendix A.1, conventional Dempster–Shafer fusion in TMVC methods satisfies

uA⊕B≤min⁡(uA,uB).u^{A\oplus B} \leq \min(u^A,u^B).

This implies that fused uncertainty monotonically decreases as more view opinions are aggregated, potentially underestimating the epistemic uncertainty of unknown samples and leading to overconfident predictions. In our frameworks, UD-PCR adaptively redistributes cross-view conflict to both beliefs and uncertainty, enabling reliable uncertainty estimation in open-set recognition. Furthermore, we derive a Jousselme consistency loss to promote cross-view epistemic coherence with linear complexity.

(3) Experimental validation.

Our ablation studies further support the complementary contributions of these components to known-class classification and unknown-class rejection.

[W4] The proposed consistency loss is not fully studied. The authors introduce a Jousselme consistency loss to improve agreement across views. However, the Jousselme distance is an existing measure, and its advantage over common losses such as MSE, KL divergence, or JS entropy is not shown. Table 3 only provides a limited ablation study. Also, the notation (LJouL_{Jou}) in Table 3 is not consistent with (LconL_{con}) in the main text.

Thank you for this constructive suggestion. We would like to clarify the motivation behind our consistency loss and provide additional comparisons with alternative objectives.

(1) Motivation for Jousselme consistency.

While the Jousselme distance is an established measure, we adapt it to cross-view subjective opinions to encourage consistency in class beliefs and uncertainty estimates. By exploiting the focal-element structure of subjective opinions, we derive a closed-form expression (Appendix A.2) that enables efficient computation with O(K)O(K) complexity.

(2) Comparison with alternative losses.

To further evaluate its effectiveness, we conducted additional experiments by replacing the Jousselme consistency loss with MSE, symmetric KL divergence, and JS divergence while keeping the remaining model components unchanged. As shown in Table 3, the Jousselme-based consistency objective achieves competitive performance on MVSA-Single, supporting its effectiveness in promoting cross-view opinion consistency and reliable open-set recognition.

Table 3. Comparison of Different Consistency Objectives on MVSA-Single.

Consistency Objective[email protected]% ↑CCR@1% ↑OSCR ↑Open AUC ↑
MSE4.70±1.887.29±2.7556.29±3.5863.09±2.01
Symmetric KL Divergence3.20±1.375.41±1.8756.89±2.7863.62±1.37
JS Divergence4.72±2.037.45±2.8656.58±2.9763.51±1.42
Jousselme-based Consistency (Ours)5.01±2.527.97±2.8456.72±2.9563.77±1.31

(3) Notation consistency.

We thank you for pointing out this inconsistency. We have carefully reviewed the entire manuscript and standardized the notation to Lcon\mathcal{L}_{con} in the revised version.

Reviewer_Ur8uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EU2O, a framework that jointly embraces aleatoric and epistemic uncertainty for trusted open-world multi-view classification. It introduces Aleatoric Uncertainty-Driven Cross-view Fusion (AU-DCF) to learn a robust joint representation, Uncertainty-Driven Proportional Conflict Redistribution (UD-PCR) to aggregate multi-view subjective opinions, and a Jousselme consistency loss to enforce cross-view epistemic consistency. Experiments on six benchmarks demonstrate improved closed-set classification and open-set recognition performance.

Critical strengths
  1. The paper addresses an important problem in trusted open-world multi-view classification by jointly embracing aleatoric and epistemic uncertainty. The motivation is clear, particularly regarding the limitations of existing TMVC methods in handling view-specific aleatoric uncertainty and maintaining epistemic coherence.

  2. The proposed UD-PCR introduces an uncertainty-driven strategy to dynamically redistribute conflict mass into belief masses and uncertainty. The paper provides theoretical analysis of the fusion operator and evaluates the proposed framework on six multi-view and multi-modal benchmarks, demonstrating improvements in closed-set classification and open-set recognition.

Critical weaknesses
  1. The novelty of the proposed framework is not sufficiently clear. AU-DCF mainly builds on Gaussian representation modeling, KL-based distribution alignment, and uncertainty-driven fusion, while KAN-Evidential networks and the Jousselme consistency loss rely on existing methods. Although UD-PCR is the main methodological contribution, the paper could provide a more detailed discussion of how it differs from existing opinion aggregation strategies.
  2. The experimental comparisons mainly focus on existing TMVC methods, without including open-world multi-view classification methods such as OpenViewer and TOMC, despite discussing them in the related work. Comparisons with these methods would help provide a more comprehensive evaluation of the proposed framework in open-set recognition.
Questions for authors
  1. Could the authors provide comparisons with open-world multi-view classification methods such as OpenViewer and TOMC under the same experimental settings?
  2. In Appendix A.5, UD-PCR shows relatively small improvements over other fusion strategies. Could the authors evaluate its effectiveness on more datasets and under different openness settings to further validate its contribution?
  3. How well does the learned Gaussian variance reflect view-specific aleatoric uncertainty? Could the authors provide further analysis of its relationship with input noise and its impact on cross-view fusion?
3 comments
AuthorsOpenReview ↗

Author Response to Reviewer Ur8u — Part 1

We sincerely thank you for your thoughtful evaluation and constructive feedback on our work. Your comments have encouraged us to further clarify the methodological contributions of EU²O, strengthen its theoretical foundations, and provide more comprehensive experimental validation. We have carefully addressed your concerns through additional analyses, experiments, and corresponding revisions. Our detailed responses are provided below.

[W1] Clarification of the Main Contributions.

We thank the reviewer for the insightful comments. We would like to clarify two key contributions.

(1) Limitations of TMVC in Open-Set Recognition.

Existing TMVC methods are primarily designed for closed-set classification, overlooking the reliability challenges introduced by unknown categories. We identify two underlying limitations: view-specific aleatoric uncertainty compromises cross-view representation alignment, while conventional opinion aggregation may suppress epistemic uncertainty essential for unknown rejection. To address these challenges, EU²O jointly models both uncertainty types at the representation and decision levels, enabling reliable classification and uncertainty estimation in open-world scenarios.

(2) Distinction from Existing Fusion Strategies.

As theoretically derived in Appendix A.1, conventional Dempster–Shafer fusion [1] satisfies

uA⊕B≤min⁡(uA,uB).u^{A\oplus B} \leq \min(u^A,u^B).

Similarly, evidence accumulation [2–4] explicitly promotes uncertainty reduction as views are integrated. While beneficial in closed-set settings, this property may suppress essential uncertainty signals from unknown samples, leading to overconfident predictions.

In our framework, UD-PCR explicitly separates orthogonal consensus from cross-view conflict and adaptively redistributes conflicting mass to both beliefs and uncertainty. This enables uncertainty recovery under strong view disagreement, addressing a critical limitation of conventional fusion strategies in open-world scenarios.

[W2/Q1] The experimental comparisons mainly focus on existing TMVC methods, without including open-world multi-view classification methods such as OpenViewer and TOMC, despite discussing them in the related work. Comparisons with these methods would help provide a more comprehensive evaluation of the proposed framework in open-set recognition.

We sincerely thank you for this valuable suggestion.

(1) Comparison with OpenViewer.

Our initial comparisons primarily focus on existing TMVC methods, as EU²O aims to achieve reliable uncertainty estimation and trustworthy decision-making. OpenViewer [5] primarily addresses unknown-category recognition by generating pseudo-unknown samples to mitigate overconfident predictions.

To provide a more comprehensive evaluation, we further include OpenViewer as an additional open-set baseline. For fairness, we adapt OpenViewer to our unified experimental protocol (e.g., identical data partitions), while preserving its original architecture, loss functions, and core learning mechanisms.

Table 1. Comparison with OpenViewer under the unified experimental protocol .

DatasetMethodOSCR ↑Open AUC ↑
Flower17OpenViewer17.33±0.6051.13±7.34
EU²O76.79±8.1879.73±5.76
VGGFace2-50OpenViewer61.33±12.5175.80±7.47
EU²O71.80±2.1979.57±1.85
YouTubeOpenViewer36.21±14.6256.79±9.12
EU²O78.93±4.6181.43±3.98
SceneOpenViewer21.81±5.7048.99±7.66
EU²O68.13±12.7871.19±10.74
MVSA-SingleOpenViewer36.09±15.8454.59±6.50
EU²O56.22±3.2463.16±1.63
NYU Depth V2OpenViewer40.51±10.3847.30±10.46
EU²O48.96±10.0153.36±9.80

As shown in Table 1, EU²O achieves higher mean OSCR and Open AUC across all six benchmarks under the unified protocol, further demonstrating its effectiveness in trusted open-set recognition.

(2) Discussion of TOMC.

Its official implementation was unavailable to us, and certain implementation details could not be reliably recovered from the paper alone. We discuss TOMC [6] and its methodological differences in the Related Work section and further clarify these distinctions in the revision.

References:

[1] Trusted Multi-View Classification, ICLR 2021.

[2] Trusted Multi-View Deep Learning with Opinion Aggregation, AAAI 2022.

[3] Reliable Conflictive Multi-View Learning, AAAI 2024.

[4] Generalized Trusted Multi-view Classification Framework with Hierarchical Opinion Aggregation, TMM 2026.

[5] OpenViewer: Openness-Aware Multi-View Learning, ACM MM 2025.

[6] Trusted Open-World Multi-View Classification with Dynamic Opinion Aggregation, ACM MM 2025.

AuthorsOpenReview ↗

Author Response to Reviewer Ur8u — Part 2

[Q2] In Appendix A.5, UD-PCR shows relatively small improvements over other fusion strategies. Could the authors evaluate its effectiveness on more datasets and under different openness settings to further validate its contribution?

We sincerely thank you for this constructive suggestion. To further validate the effectiveness and generalizability of UD-PCR, we extend the fusion strategy comparison to Scene and YouTube under openness={0.3,0.4}openness = \{0.3, 0.4\}, while keeping the remaining experimental settings unchanged.

As shown in Table 2, UD-PCR achieves the highest OSCR and Open AUC across all four settings, demonstrating its effectiveness in open-set recognition. Furthermore, Table 3 shows that UD-PCR achieves the best overall average rank of 1.78 (final rank: 1) across all evaluated metrics and settings.

Table 2. Comparison of different fusion strategies on Scene and YouTube under different openness settings.

DatasetOpennessFusion[email protected]% ↑CCR@1% ↑OSCR ↑Open AUC ↑Avg. Rank ↓
Scene0.3MEAN5.01±4.288.81±6.0461.12±14.0265.11±12.493.50
TMC4.84±3.808.17±4.9661.57±13.7065.43±12.113.75
RCML5.13±3.708.52±5.8361.01±13.1565.49±11.142.75
TMCEK5.03±3.669.08±6.5560.99±13.4065.35±11.473.25
UD-PCR5.38±3.848.45±5.0461.96±13.8365.81±12.201.75
0.4MEAN7.51±7.0910.55±8.6959.88±8.3962.50±8.204.75
TMC7.56±6.5310.66±8.7960.64±8.1863.14±8.112.50
RCML8.17±6.5911.37±8.9160.34±8.2863.11±7.882.00
TMCEK7.81±6.3310.75±8.3260.32±8.2163.08±7.793.00
UD-PCR7.48±6.4510.64±8.5061.46±7.9464.04±7.682.75
YouTube0.3MEAN8.83±5.3914.83±8.9068.36±8.5474.50±6.412.25
TMC8.67±5.0215.33±8.3167.11±10.0873.70±7.352.75
RCML6.33±7.0210.50±10.8262.09±10.9968.25±8.864.00
TMCEK5.67±6.6510.00±10.2261.96±10.7968.16±8.705.00
UD-PCR9.67±4.5516.33±11.0468.75±8.5674.65±6.601.00
0.4MEAN2.25±1.663.75±3.2659.65±9.4662.33±10.291.50
TMC1.75±1.003.25±2.3258.21±9.2861.27±9.982.88
RCML0.50±0.610.75±0.6151.27±10.3154.19±11.764.88
TMCEK0.50±0.611.00±0.9451.52±9.9654.47±11.464.13
UD-PCR2.00±1.273.25±2.8160.75±8.8163.45±9.501.63

Table 3. Average ranking of fusion strategies across datasets and openness levels (lower is better).

DatasetOpennessMEANTMCRCMLTMCEKUD-PCR
Scene0.33.503.752.753.251.75
0.44.752.502.003.002.75
YouTube0.32.252.754.005.001.00
0.41.502.884.884.131.63
Overall Avg. Rank ↓—3.002.973.413.841.78
Final Rank ↓—32451

These additional results further support the effectiveness of UD-PCR. The extended comparisons have also been incorporated into the revised manuscript.

AuthorsOpenReview ↗

Author Response to Reviewer Ur8u— Part 3

[Q3] How well does the learned Gaussian variance reflect view-specific aleatoric uncertainty? Could the authors provide further analysis of its relationship with input noise and its impact on cross-view fusion?

Thank you for this insightful question. We provide the following clarifications and additional experimental analysis.

(1) Regarding aleatoric uncertainty modeling.

Our approach follows established probabilistic representation learning methods (Shi and Jain, ICCV 2019 [1]; Tang et al., ACL 2025 [2]) and recent distribution-based aleatoric uncertainty modeling (Tang et al., IJCAI 2025 [3]; Wang et al., CVPR 2026 [4]).

Specifically, in our framework, each sample is mapped by encoders into a mean vector and a standard deviation vector, which jointly parameterize its latent Gaussian distribution. The mean captures view-specific semantic information, while the variance characterizes uncertainty associated with noisy or ambiguous observations, providing a principled basis for modeling view-specific aleatoric uncertainty. During optimization, we align cross-view Gaussian distributions and regularize them toward a standard normal prior to suppress redundant information and improve uncertainty estimation.

(2) Regarding the relationship with input noise and cross-view fusion.

We conducted additional experiments on the Scene dataset by injecting Gaussian noise of varying intensities into last view (View 3). The corresponding experimental results are provided below:

As shown in Fig. 1(a), the change in predicted variance exhibits a positive Pearson correlation with Gaussian noise intensity (r=0.214r=0.214). Fig. 1(b) further shows that the predicted variance increases as the corrupted view becomes noisier, while the uncorrupted views remain relatively stable. These observations suggest that the learned Gaussian variance is sensitive to view-specific input variability, supporting its role as an estimator of aleatoric uncertainty.

Furthermore, Fig. 1(c) shows that AU-DCF correspondingly reduces the contribution of the corrupted view, while Fig. 1(d) demonstrates better Open AUC than uniform fusion under increasing noise.

These results provide empirical support that the learned Gaussian variance reflects aleatoric uncertainty inherent in the data, enabling AU-DCF to adaptively mitigate the influence of unreliable views and improve the robustness of cross-view fusion.

References:

[1] Probabilistic Face Embeddings, ICCV 2019.

[2] Modeling Uncertainty in Composed Image Retrieval via Probabilistic Embeddings, ACL 2025.

[3] DFMU: Distribution-based Framework for Modeling Aleatoric Uncertainty in Multimodal Sentiment Analysis, IJCAI 2025.

[4] Uncertainty-Aware Modality Fusion for Unaligned RGB-T Salient Object Detection, CVPR 2026.


We hope our response could address your concerns, and we also welcome any further feedback you might provide to enhance our manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear reviewers and AC,

We sincerely thank you for your valuable time and constructive feedback, which have helped us strengthen our manuscript. We have carefully considered all the concerns and suggestions and made corresponding revisions through additional experiments, theoretical clarifications, and methodological improvements. The main updates are summarized as follows.

1. Additional experiments:

  • We conducted additional robustness experiments under varying Gaussian noise intensities and conflict ratios, evaluating both closed-set classification and open-set recognition.
  • We evaluated predictive calibration using ECE, NLL, and Brier Score across five datasets.
  • We extended the UD-PCR fusion strategy comparisons to Scene and YouTube under different openness settings and conducted a controlled KAN-to-MLP ablation.
  • We included comparisons with OpenViewer under a unified experimental protocol across six benchmarks.
  • We compared the Jousselme consistency objective with MSE, symmetric KL divergence, and JS divergence.

2. Theoretical clarifications:

  • We added finite-evidence analysis to characterize the relationship between cross-view conflict and fused uncertainty.
  • We clarified the theoretical scope of UD-PCR under sequential multi-view fusion.
  • We addressed the boundary case involving two vacuous opinions and revised the corresponding theoretical statement.

3. Methodological clarifications:

  • We further clarified the distinction between EU²O and existing uncertainty-aware multi-view learning and opinion aggregation methods.
  • We clarified the known/unknown class partitioning.
  • We provided additional explanations of Gaussian variance modeling and its relationship with input noise and adaptive cross-view fusion.

In summary, our rebuttal addresses every concern raised by the three reviewers through additional experiments and detailed mechanistic analyses. In the revised manuscript, all newly added and modified contents are highlighted in blue for ease of inspection. We sincerely hope these revisions address your concerns and provide a clearer presentation of the contributions and effectiveness of EU²O.

We wish you a pleasant day!


Additionally, we appreciate the contributions of the SACs and ACs in coordinating the discussion and evaluation.

Sincerely,

The Authors of Submission 7946.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.