acceptodds
Withdrawn from ICLR 2027

ReLoF: Reliability-Guided Logarithmic Fusion for Robust Multi-View Classification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-view learning integrates heterogeneous observations of the same object to improve classification modeling. Despite numerous advances, robustness remains a persistent challenge due to the presence of low-quality or conflicting views. Effectively modeling view reliability in the fused space is therefore nontrivial for robust multi-view classification. This paper proposes Reliability-Guided Logarithmic Fusion (ReLoF), which unifies instance-wise reliability estimation and the logarithmic opinion pool (LogOP) within a categorical probability-opinion framework. In a nutshell, ReLoF first forms a categorical probability opinion for each view, and then estimates its reliability using normalized-entropy confidence and Jensen–Shannon-based inter-view consistency relative to a leave-one-out consensus. Their normalized values serve as LogOP exponents for aggregating class evidence. In addition, ReLoF uses Maximum-Logit (MaxLog) regularization to moderate excessively sharp single-view predictions and mitigate the effect of near-zero true-class probabilities assigned by erroneous, overconfident views. Experimental results on seven benchmark datasets show that ReLoF outperforms baselines remarkably under both normal and conflictive conditions. Ablation studies further show that LogOP improves over conventional weighted averaging and that MaxLog yields consistent numerical improvements under view conflict.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-view Learning, Multi-view Classification, View Reliability

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_T8BAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes ReLoF, which estimates sample-level reliability through entropy-based confidence and leave-one-out view consistency, and uses the reliability as the exponent of LogOP for multi-view fusion. In addition, a MaxLog regularization is introduced to alleviate the soft veto induced by erroneously high-confidence views. Experiments cover both normal and conflict scenarios on seven datasets.

Critical strengths
  1. The method is concise and interpretable, connecting view reliability, fusion geometry, and confidence control in an easy-to-implement framework.
  2. The AOP–LogOP comparison is carefully designed, with results reported on seven datasets over ten random seeds.
Critical weaknesses
  1. The protocol may leaks test information: without a validation set, the model is evaluated on the test set every 10 epochs with the best test accuracy retained, and normalization precedes the data split; given gains as small as 0.07–0.30 percentage points, this may alter the main conclusions.
  2. The core reliability estimator lacks ablations against uniform-weighting, entropy-only, and consistency-only variants.
  3. Only one view is replaced while most views remain clean, matching its prior assumption, so robustness under general conflicts is unproven.
Questions for authors
  1. Could the authors use an independent validation set for model selection, normalize using training-set statistics only, and re-report all results?
  2. Could comparisons with uniform-weighting, confidence-only, consistency-only, and other confidence regularization methods be added?
  3. How do ReLoF's performance and reliability weights change as the number of conflicting views grows from one to a majority?
Reviewer_FW8GOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Aiming at reliable multi-view classification, this paper models the reliability of each view by intra-view confidence and inter-view consistency. With view-wise weights, this paper implements fusion as a softmax over weighted logits. The authors provide theoretical insights and extensive experiments on seven multi-view benchmarks, demonstrating consistent improvements over state-of-the-art trusted multi-view methods under both normal and conflictive test settings.

Critical strengths
  1. Elegant Theory with Practical Implementation The proof that LogOP is strictly equivalent to applying softmax over reliability-weighted logits is mathematically elegant yet highly practical, entirely avoiding numerical underflow to ensure stable, efficient end-to-end training.
Critical weaknesses
  1. Limited Methodological Novelty The framework primarily assembles existing techniques (LogOP, entropy, JS divergence, and MaxSup-style regularization), offering moderate rather than fundamental algorithmic innovation.

  2. Heuristic Reliability Formulation The multiplicative reliability score is an ad hoc design that lacks formal theoretical guarantees, optimization bounds, or deeper justification for its specific functional form.

  3. Fragility of Leave-One-Out Consensus The reliability audit inherently assumes a reliable majority. Under correlated view degradation or when erroneous views form a majority, it will falsely penalize the correct minority view.

  4. Synthetic Conflictive-Test Settings The simulated view degradation relies on independent Gaussian noise and cross-class feature replacement. This lacks realism compared to structured, correlated real-world failures.

  5. Insufficient Comparison with EDL While Evidential Deep Learning methods share similar challenges, the paper lacks an in-depth conceptual and empirical comparison with Dirichlet-based fusion approaches.

Questions for authors
  1. Could you provide more theoretical justification for the specific multiplicative form (Eq. 4)? Why was this chosen over alternative formulations (e.g., additive or learned attention), and how sensitive is the model to the temperature?

  2. The current consensus mechanism inherently assumes that the majority of views are reliable. How does ReLoF behave when multiple views suffer from correlated degradation and consistently provide erroneous evidence? Is there a mechanism to detect or mitigate this "majority-bias" failure mode?

  3. The evaluation simulates view degradation using independent Gaussian noise and cross-class feature replacement. In real-world scenarios, sensor failures are often structured and correlated. Could you discuss whether the framework was tested under more realistic, structured corruption patterns, and how it would perform?

  4. Since Dempster's rule in EDL methods (e.g., TMC, ETMC) is also vulnerable to Zadeh's paradox (a form of soft-veto), could you elaborate on the fundamental conceptual differences between ReLoF's reliability-weighted LogOP and EDL-based evidence aggregation? Would it be possible to provide a direct empirical comparison under the same conflictive settings?

  5. The paper mentions that the reliability score is detached from the computational graph during training. Does this detachment prevent the private encoders from learning representations that are inherently more "trustworthy" or robust to conflicts? Did you experiment with backpropagating through the reliability module?

Reviewer_7Cf6OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Reliability-Guided Logarithmic Fusion (ReLoF) for robust multi-view classification. Each view produces a categorical opinion, whose reliability is estimated by combining entropy-based confidence and Jensen–Shannon consistency against a leave-one-out consensus. The resulting reliability weights are used for logarithmic opinion pooling, implemented through weighted logits. Maximum-logit regularization is further introduced to mitigate overconfident predictions. Experiments on seven datasets evaluate the method under normal and synthetic conflict conditions, with additional pooling and regularization analyses.

Critical strengths
  • The reliability estimator combines predictive confidence with agreement against a leave-one-out consensus, providing a reasonable mechanism for reducing the influence of conflicting views during fusion.

  • The weighted-logit formulation enables efficient logarithmic opinion pooling. The experiments cover seven datasets under normal and conflictive conditions, with controlled pooling comparisons and parameter sensitivity analyses.

Critical weaknesses
  • Appendix B.1 states that the best test accuracy is selected from evaluations conducted every 10 epochs, without a separate validation set. This introduces test-set selection bias and compromises the reliability of the reported comparisons. Moreover, Min-Max normalization is performed before the train-test split, potentially leaking test-set statistics into training.

  • The reliability estimator is not compared with uniform, confidence-only, or consistency-only weighting under the same LogOP framework. The existing pooling ablations therefore do not establish whether combining confidence and inter-view consistency is necessary.

  • The accuracy improvements from MaxLog in Table 5 are relatively small, ranging from 0.06 to 0.59 percentage points. The experiments do not directly demonstrate whether MaxLog mitigates the claimed soft-veto effect caused by confidently incorrect views rather than providing general regularization benefits.

  • The leave-one-out consistency estimator relies on agreement among other views to assess reliability. When multiple views share incorrect predictions, this mechanism may suppress a correct but inconsistent view. The experiments do not systematically examine how performance changes with the number or proportion of conflicting views.

Questions for authors
  • Why is the best test accuracy used for checkpoint selection rather than a separate validation set? Can the authors report results using a validation-based selection protocol with preprocessing fitted only on training data?

  • Can the authors provide controlled comparisons using uniform, confidence-only, consistency-only, and combined reliability weights under the same LogOP framework?

  • Can the authors provide direct evidence that MaxLog reduces erroneous-view logit dominance or the soft-veto effect, beyond the reported accuracy improvements?

  • How does ReLoF perform when multiple views provide consistent but incorrect predictions, particularly when the reliable views form a minority?

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

Related papers

Discussion (0)

Sign in to comment.