acceptodds
Under review as a conference paper at ICLR 2027

Robust Multi-View Learning via Hybrid Geometric Reliability under Test-Time Noisy Correspondence

Abstract

The effectiveness of multi-view learning hinges on correct correspondences across different views. However, data acquisition and transmission failures in practical deployment may cause certain views of a sample to be mismatched, leading to test-time noise correspondence that severely undermines multi-view fusion. Existing methods predominantly rely on view-wise statistical information for reliability assessment, while largely overlooking the intrinsic geometric structure of multi-view features. To address these limitations, we propose a robust multi-view learning framework based on Hybrid Geometric Reliability, termed HGR. First, Noisy Correspondence Augmented Learning simulates diverse view mismatch patterns during training and learns view reliability through explicit supervision. Second, Hybrid Geometric Learning constructs Euclidean and Hyperbolic Gramian matrices to capture the high-order joint geometric structure of multiple views, and performs reliability evaluation based on cross-view structural relationships. Furthermore, the Geometric Consistency Principle is introduced, where the Gramian volume contrastive loss and the view consistency loss are devised. The two losses reinforce the geometric coherence of correctly matched samples while amplifying the structural distortion of mismatched ones. Extensive experiments on multiple multi-view benchmark datasets demonstrate that the proposed HGR achieves superior performance and robustness over the state-of-the-art methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Robust Multi-View Learning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_1cHWOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a robust multi-view learning method based on a hybrid geometric reliability mechanism. The proposed method contains two modules, noisy correspondence augmented learning is introduced to generate the mismathced samples and supervise the training of the reliability estimator, hybrid geometric learning is introduced to maximize the geometric agreement across views. The experimental results on 12 MV datasets show that the proposed method is superior over the SOTA methods.

Critical strengths
  1. Different from previous methods that rely on pairwise statistics to determine view reliability, this method proposes a view reliability discrimination approach based on the overall geometry of views. It constructs the geometric structure of views jointly from Euclidean Gram geometry and Hyperbolic Gram geometry, fully exploiting the complementarity between Euclidean and hyperbolic geometries. This practice provides unique insights for the field.
  2. A contrastive loss based on Gram volume is proposed, which aims to minimize the volume of view-aligned samples while maximizing the volume of view-unaligned samples from a geometric perspective, thereby effectively separating aligned and unaligned samples. Furthermore, a view consistency loss is introduced to make the volume loss more directional.
Critical weaknesses
  1. The explanation of Figure 1 is insufficient. First, it does not clarify why the reliability densities output by BML overlap extensively, what such overlap implies, and what causes it. Second, it fails to specify what the box plots at the top of the figure represent and what important phenomenon they reveal. These missing details directly hinder the understanding of the HGR's advantages.
  2. The authors construct the Gram geometry of samples from both Euclidean and hyperbolic spaces, which raises the question of whether this is redundant. Since the downstream task is classification, which emphasizes the separability of different classes, it is unclear whether mining fine-grained hierarchical structure is meaningful. The authors should explain why hyperbolic geometry is further adopted.
  3. In Equation 21, the final Gram volume is obtained by fusing the Euclidean volume and the hyperbolic volume, with beta as a trainable weight parameter. The authors should provide the specific value that beta converges to when the algorithm reaches convergence, so that the necessity of the hyperbolic volume can be directly judged.
Questions for authors
  1. The authors introduce hyperbolic Gram geometry as a complement to Euclidean Gram geometry, but its necessity is not well justified. The authors should provide a proper argument for why hyperbolic space is needed.
  2. The value of beta at convergence should be reported, as it reflects how the Euclidean Gram volume and the hyperbolic Gram volume are balanced.
1 comment
AuthorsOpenReview ↗

Response to Reviewer 1cHW

We thank the reviewer for the constructive comments. We have responded to these questions one by one.

Response to W1: Explanation of Figure 1

The substantial overlap between the reliability distributions of clean and noisy views in BML primarily arises from its reliance on prediction-level statistics, including view-wise entropy and cross-view prediction consistency, which may fail to adequately capture structural distortions caused by view mismatches. Such overlap indicates that BML struggles to distinguish reliable from unreliable views, potentially assigning high reliability scores to corrupted views and misleading the final prediction. In contrast, HGR exploits Euclidean and Hyperbolic Gramian geometry to capture cross-view structural relationships, resulting in better separation between clean and noisy views. Furthermore, the box plots at the top of Figure 1 summarize the reliability score distributions, including their medians, interquartile ranges, and overall dispersion. Compared with BML, HGR exhibits a clearer separation between the two groups, demonstrating its stronger reliability discrimination capability.

Response to W2: Necessity of Hyperbolic Geometry

We emphasize that hyperbolic geometry is introduced not to model hierarchical class structures, but to enhance view reliability assessment through complementary geometric information. Specifically, Euclidean Gramian geometry captures cross-view relationships through conventional inner products, while Hyperbolic Gramian geometry employs Lorentzian inner products to characterize structural relationships from a different geometric perspective. Their combination provides richer geometric representations for distinguishing reliable and unreliable views. Importantly, hybrid geometry is employed for view reliability assessment rather than direct classification, and thus does not require explicit class hierarchies. Furthermore, the following ablation results in Table 1 demonstrate that combining both geometries achieves better performance than using either alone, validating their complementarity.

Table 1. Ablation study with respect to Euclidean Gramian geometry(EGG) and Hyperbolic Gramian geometry(HGG).

EGGHGGGRAZ02Caltech8677flower17MITIndoorNUSWIDEOBJAWA
✗✓67.16±1.8085.20±0.6486.25±0.9557.05±1.0247.00±0.5241.77±0.27
✓✗67.57±1.5985.40±0.3286.03±1.2359.57±0.9747.53±0.5342.96±0.18
✓✓68.58±2.9185.73±0.5186.47±1.3459.89±1.4247.57±0.2243.15±0.27

Response to W3: Analysis of the Learnable Fusion Weight β\beta

In Eq. (21), the learnable parameter β\beta adaptively balances the contributions of Euclidean and Hyperbolic Gramian volumes. To clarify their relative importance, we further examine the converged values of β\beta across different datasets. The results in Table 2 show that β\beta converges to [0.1, 0.8], indicating that both geometric volumes contribute to the optimization. These observations further support the effectiveness of hybrid geometric modeling.

Table 2. The converged valus of β\beta on six multi-view datasets.

DatasetGRAZ02Caltech8677flower17MITIndoor100LeavesWikipedia
β\beta0.360.110.500.790.350.33
Reviewer_rjMnOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper handles the challenge of test-time noise correspondence in mvl, proposing the HGR, a robust mvl framework to overcome the limitations of existing methods that neglect the intrinsic geometric structure of mv features. Extensive experiments on large-scale multi-view datasets demonstrate that HGR effectively distinguishes clean views from noisy ones, achieving superior classification performance and noise robustness compared to existing algorithms.

Critical strengths
  1. This work has a clear motivation. The authors illustrate the flaws of existing methods, noting that pairwise‑statistics‑based view reliability estimation cannot capture global inter‑view correlations.

  2. The authors propose the HGR framework, which leverages view‑level geometry to judge view reliability and designs a set of geometry‑guided losses to achieve state‑of‑the‑art view reliability discrimination.

  3. Extensive comparative and ablation experiments are conducted on multiple multi‑view datasets. The experimental results demonstrate the effectiveness of the proposed geometry‑based view discrimination and cross‑view consistency enhancement.

Critical weaknesses
  1. Both BML and the proposed HGR rely on constructing noise correlated samples in the training set to train the perspective reliability evaluator. Currently, the perturbation ratios of both seem to be set at fixed values in the experiments, and the impact of different training perturbation ratios on the performance has not been analyzed. In particular, the effectiveness of the training mechanism should be verified under low perturbation ratios.

  2. It is suggested that the discussion of related work be expanded, particularly to include existing trustworthy multi-view classification methods based on EDL, such as [1], among others, to provide a more comprehensive review of the relevant literature and better contextualize the proposed method. Moreover, only one of the current compared methods is from 2026. Given the rapid advances in multi‑view learning, comparing against merely one recent baseline cannot fully validate the efficacy of the proposed algorithm. More state‑of‑the‑art multi‑view learning approaches should be added for comparison.

  3. This algorithm involves computations for numerous modules, such as evaluator training and Gram matrix calculation, which are computationally expensive. Since runtime serves as a vital metric for algorithm evaluation, the authors should report corresponding runtime results, especially on large‑scale datasets.

[1] Enhancing testing-time robustness for trusted multi-view classification in the wild.

Questions for authors
  1. Does the perturbation ratio of noisy correspondences in the training set exert a significant impact on experimental performance? Specifically, whether the training mechanism remains valid under low‑perturbation settings?

  2. The discussion and comparison against recent multi‑view learning baselines is insufficient. More state‑of‑the‑art recent approaches should be included for evaluation.

  3. The runtime of compared baselines should be reported, especially on large‑scale datasets, computational cost is a key practical metric for multi‑view learning algorithms.

2 comments
AuthorsOpenReview ↗

Response to Reviewer rjMn (1/2)

We thank the reviewer for their valuable comments, and we have carefully responded to each comment as following.

Response to W1: Impact of Training Perturbation Ratios

To investigate the impact of training perturbations, we conduct additional experiments with perturbation ratios of 0.1, 0.2, 0.3, 0.4, and 0.5. The results in Tables 1, 2, 3, 4, 5, 6 show that reducing the training perturbation ratio leads to some performance degradation, as fewer corrupted samples provide insufficient supervision for learning view reliability. Nevertheless, HGR consistently outperforms competing methods even under low perturbation ratios, demonstrating the effectiveness and robustness of the proposed training mechanism.

Table 1. Performance with different training perturbation ratios on GRAZ02.

Method0.10.20.30.40.5
BML63.5163.5166.5566.5567.23
HGR65.8867.9168.2469.5968.58

Table 2. Performance with different training perturbation ratios on Caltech8677.

Method0.10.20.30.40.5
BML83.784.0484.9184.8584.26
HGR85.0886.0086.0686.0685.73

Table 3. Performance with different training perturbation ratios on flower17.

Method0.10.20.30.40.5
BML80.8883.4683.8284.9384.63
HGR83.0983.4686.0386.7686.47

Table 4. Performance with different training perturbation ratios on MITIndoor.

Method0.10.20.30.40.5
BML51.5953.1754.5753.6455.24
HGR57.2858.4958.6859.6159.89

Table 5. Performance with different training perturbation ratios on 100Leaves.

Method0.10.20.30.40.5
BML94.6995.0095.9495.6296.16
HGR94.6995.3196.5696.8896.75

Table 6. Performance with different training perturbation ratios on Wikipedia.

Method0.10.20.30.40.5
BML51.851.0849.6453.2453.38
HGR53.2452.5254.6853.2454.68

Response to W2: Expanded Related Work and Comparisons with Recent SOTA Method

  1. We have added a discussion on EDL-based trustworthy multi-view learning in the relevant work (added to the manuscript), with the specific content as follows: Evidential Deep Learning (EDL) has been widely explored in trustworthy multi-view learning to quantify predictive uncertainty and improve fusion reliability. The studies [1,2] were the first to introduce evidential deep learning into multi-view learning, achieving trustworthy fusion of multiple views based on the Dempster-Shafer theory. Building upon this framework, subsequent studies [3,4,5,6,7,8] have explored various evidence aggregation strategies to further enhance fusion reliability. These methods primarily focus on estimating view-specific reliability to guide opinion aggregation, and improving cross-view information interaction through hierarchical or adaptive fusion mechanisms. Recent study [9] has further improved robustness under test-time view corruptions by identifying and down-weighting unreliable evidence through statistical hypothesis testing. While these methods provide valuable theoretical insights and demonstrate promising empirical performance, they remain ineffective in addressing test-time noisy correspondence, as they primarily focus on uncertainty estimation and evidence aggregation without explicitly modeling cross-view correspondence reliability.

[1] Trusted multi-view classification, ICLR, 2021.

[2] Trusted multi-view classification with dynamic evidential fusion, IEEE TPAMI, 2023

[3] Trusted multi-view deep learning with opinion aggregation, AAAI, 2022.

[4] Reliable conflictive multi-view learning, AAAI, 2024.

[5] Navigating conflicting views: Harnessing trust for learning, ICML, 2025.

[6] Trusted multi-view classification via evolutionary multi-view fusion, ICLR, 2025.

[7] Generalized trusted multi-view classification framework with hierarchical opinion aggregation, IEEE TMM, 2026.

[8] Beyond magnitude: Scale-invariant evidential fusion for multi-view classification, ICML, 2026.

[9] Enhancing testing-time robustness for trusted multi-view classification in the wild, CVPR, 2025.

AuthorsOpenReview ↗

Response to Reviewer rjMn (2/2)

  1. We have incorporated an additional SOTA method UVF published in AAAI 2026 and conducted comprehensive comparisons across 12 datasets under the same experimental settings. The results in Tables 7, 8 demonstrate that our proposed HGR consistently outperforms this recent baseline, further validating its effectiveness and competitiveness against the latest advances in multi-view learning.

Table 7. Performance comparison on six datasets, where NC rates are set to 0.5 and 1.0.

MethodGRAZ (0.5)GRAZ (1.0)Caltech (0.5)Caltech (1.0)flower (0.5)flower (1.0)MIT (0.5)MIT (1.0)Leaves (0.5)Leaves (1.0)Wiki (0.5)Wiki (1.0)
UVF61.4951.3570.9148.7375.7463.2451.4933.1284.3865.9453.9630.22
BML67.2359.0684.2677.2184.6376.9155.2442.8996.1693.5653.3835.68
HGR68.5861.4985.7380.3786.4779.0459.8949.7796.7594.1354.6837.70

Table 8. Performance comparison on six datasets, where NC rates are set to 0.5 and 1.0.

MethodLandUse (0.5)LandUse (1.0)MNIST (0.5)MNIST (1.0)CCV (0.5)CCV (1.0)NUSWIDE (0.5)NUSWIDE (1.0)AwAfea (0.5)AwAfea (1.0)YouTube (0.5)YouTube (1.0)
UVF57.1442.6285.8077.4043.8434.3234.0724.1529.9617.2968.8553.60
BML75.6269.0296.4395.7151.7344.1645.9939.9840.5232.2284.1278.42
HGR76.8672.4396.9396.4853.4946.8347.5743.1543.1537.2286.5383.39

Response to W3: Comparison of Running Time

We have evaluated the runtime of all compared methods across all datasets under identical hardware configurations in Tables 9, 10. Although HGR does not achieve the shortest runtime, it remains computationally efficient compared with several existing methods while consistently achieving superior classification performance.

Table 9. Comparison of running time (seconds) on six datasets.

MethodGRAZ02Caltech8677flower17MITIndoor100LeavesWikipedia
TMC18.8495.8123.0957.5457.824.28
UIMC33.09192.0342.59108.5625.198.44
ECML23.6345.9529.6292.1417.034.26
CCML18.70101.6722.9661.6311.554.10
MAMC8.72214.0420.25137.9837.794.76
ETF76.69504.07115.59293.4865.7619.20
TMCEK100.04503.89336.16526.5184.7716.98
FUML60.32778.24128.85533.6726.869.79
RML711.24694.23871.21864.27403.33213.28
BML13.5066.6717.6439.416.101.82
HGR20.02151.5731.5899.0311.994.34

Table 10. Comparison of running time (seconds) on six datasets.

MethodLandUseMNIST10kCCVNUSWIDEOBJAWAYouTubeFace
TMC15.2763.5345.20295.95438.701180.60
UIMC33.47161.29106.88650.881142.452668.09
ECML15.5390.4571.70476.86628.071246.56
CCML20.8876.2554.98322.30454.221109.86
MAMC20.8354.4247.74423.96720.611357.66
ETF96.11434.17287.011703.692262.305112.54
TMCEK214.041054.48818.445439.115916.235823.10
FUML44.26181.25137.41560.562265.533555.63
RML463.60461.67501.283169.68722.332493.28
BML7.6634.3323.34147.12414.51667.45
HGR20.2992.4767.35428.271257.561936.03
Reviewer_ehXROpenReview ↗
Rating
4(+1)
Confidence
3(+2)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors address the problem of view misalignment that frequently occurs in multi-view data at test time by proposing a robust learning framework called HGR. It first trains an classifier to identify unreliable views by simulating misalignment patterns, and then uses Gramian geometry in both Euclidean and Hyperbolic spaces to assess the reliability of each view from a holistic structural perspective. Clean views are assigned high scores, while misaligned views are suppressed. Meanwhile, the authors design two geometric consistency losses to further widen the structural discrepancy between clean and misaligned samples in the feature space. Experiments on multiple multi-view datasets show that HGR outperforms existing methods across various noise ratios, and ablation studies verify the necessity of each loss term and of the dual-geometry combination.

Critical strengths

S1. This paper is written fluently and logically, so readers can quickly grasp the motivation of the work and the main innovations of the method. The introduction clearly identifies the limitations of existing approaches, and the method section presents each component in a well-organized way. S2. The authors characterize cross-view correspondence from a geometric perspective and design a set of rules that maximize the geometric consistency of aligned samples while enforcing geometric dispersion for misaligned ones, which offers a novel viewpoint for addressing the multi-view alignment problem. S3. To verify the effectiveness of the proposed algorithm, the authors conduct experiments on 12 multi-view datasets, and the experimental results show that the proposed method achieves good performance under various noise ratios.

Critical weaknesses

W1: The authors do not analyze the computational complexity of the algorithm, whereas computational complexity is important for measuring the scalability of an algorithm.

W2: In Equation 22, the authors define a volume-based contrastive learning mechanism that aims to minimize the volume of aligned samples while maximizing that of unaligned samples, which already constitutes a strong constraint for separating aligned from unaligned samples. However, the authors then introduce an additional view consistency loss, which appears to be redundant, and the necessity of this loss should be clearly explained.

W3: Table 5 records the experimental results of all algorithms on clean multi-view data, yet the proposed method fails to achieve the best results on some datasets. The specific reasons for this should be clarified, and it should further be examined whether this implies that the algorithm actually has no advantage under mild noise contamination.

W4: From Figure 3, a smaller lambdaclambda_c is more beneficial for performance improvement. Does this suggest that the LconL_con loss actually provides little benefit to model training, or that this value must exceed a certain threshold before it produces a positive effect on training?

W5: In the test phase, the authors restrict the number of noisy views to no more than half of the total number of views. However, in real-world scenarios, more extreme cases may still occur. The effectiveness of the proposed method cannot be guaranteed when the number of noisy views exceeds half of the total number of views.

Questions for authors

Please refer to the Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Review ehXR (1/2)

We are very grateful to the reviewer for their professional and detailed comments, which are very important for improving the quality of the paper. Below we provide a point-by-point response to the reviewer's questions.

Response to W1: Computational Complexity Analysis

The computational complexity of HGR primarily arises from three components: view-specific encoding, Gramian matrix construction, and Gramian volume loss computation. First, Given the dimensions of the vv-th encoder as dvd^v, l1l_1, l2l_2, ... lll_l, CC, for the multi-view encoding process, its computational complexity is O((∑v=1Vd(v)l1+V(l1l2+⋯+ltC))N)O\left(\left(\sum_{v=1}^{V}d^{(v)}l_1+V(l_1l_2+\cdots+l_tC)\right)N\right), where NN and VV denote the numbers of samples and views. Let lmax⁡=max⁡(∑v=1Vd(v)l1,  Vl1l2,  …,  VltC)l_{\max}=\max\left(\sum_{v=1}^{V}d^{(v)}l_1,\;Vl_1l_2,\;\ldots,\;Vl_tC\right), then the encoding complexity is simplified as O(lmaxN)\mathcal{O}(l_{max}N). Second, constructing the Euclidean and Hyperbolic Gramian matrices involves pairwise inner products among VV view embeddings, resulting in a complexity of O(NV2d)O(NV^2d), where dd denotes the dimension of latent embedding. Third, computing the Gramiam volume loss takes O(Na(V2d+NuV2d))O(N_a(V^2d+N_uV^2d)), where NaN_a and NuN_u denote the numbers of aligned samples and unaligned samples, respectively. Overall, the computational complexity of the proposed HGR is summarized as O(lmaxN+Na(V2d+NuV2d))\mathcal{O}(l_{max}N+ N_a(V^2d+N_uV^2d)).

Response to W2: Necessity of the View Consistency Loss

Although both losses encourage geometric consistency, they impose complementary constraints. Specifically, LvolL_{\mathrm{vol}} provides a global geometric constraint by distinguishing aligned from unaligned samples through Gramian volume. However, a small Gramian volume does not necessarily imply strong cross-view consistency, as linearly dependent features can yield zero volume without being directionally aligned. In contrast, LconL_{\mathrm{con}} provides an explicit view-level consistency constraint by maximizing the cosine similarity between each view embedding and their mean representation, thereby offering more direct optimization guidance. The ablation results in the Table 3 of manuscript further support this complementarity.

Response to W3: Performance on Mild Noise

HGR is primarily designed to address test-time NC rather than maximize classification accuracy on clean data. When all views are correctly aligned, reliability assessment provides limited additional benefits, while the geometric regularization introduces constraints that are not always optimal for clean data. Nevertheless, this does not necessarily imply limited robustness under mild noise, where identifying and suppressing unreliable views remains beneficial. To further investigate this issue, we report additional experiments under mild noise ratios (10%, 20%, and 30%) on Caltech8677, MITIndoor, and CCV in Tables 1, 2, 3, where HGR does not achieve the best results. It can be observed that HGR still maintains superior in most cases.

Table 1. Performance comparison under NC rate 0.1.

MethodCaltech8677MITIndoorCCV
TMC89.7562.6443.21
UIMC90.8868.7844.30
ECMI88.6059.1741.38
CCML88.9062.6742.30
MAMC68.0337.2352.70
ETF91.4172.0746.69
TMCEK89.6660.5041.64
FUML81.7558.9946.20
RML83.5969.1446.28
BML90.1064.2758.23
HGR90.4666.158.25

Table 2. Performance comparison under NC rate 0.2.

MethodCaltech8677MITIndoorCCV
TMC85.4959.5042.23
UIMC86.7465.1343.06
ECML84.0156.4640.16
CCML84.5359.2441.14
MAMC65.7734.9151.32
ETF87.3468.3045.02
TMCEK85.1757.0540.84
FUML80.0555.9444.82
RML80.0065.3945.01
BML88.2761.9456.84
HGR89.0464.1157.12

Table 3. Performance comparison under NC rate 0.3.

MethodCaltech8677MITIndoorCCV
TMC80.6156.2341.19
UIMC82.2861.2941.90
ECML78.9953.6238.97
CCML79.7955.8139.80
MAMC63.7032.3649.72
ETF83.0464.3544.02
TMCEK80.3953.7539.96
FUML77.9852.9344.22
RML75.4261.5244.00
BML86.8959.6855.47
HGR87.9762.4156.32
AuthorsOpenReview ↗

Response to Reviewer ehXR (2/2)

Response to W4: Effectiveness of the View Consistency Loss

To further investigate the effectiveness of LconL_{\mathrm{con}}, we conduct additional experiments on GRAZ02 and Caltech8677 with λc∈0,0.0001,0.001,0.01,0.05,0.1\lambda_c \in {0, 0.0001, 0.001, 0.01, 0.05, 0.1}. The results in Table 4 show that even a small weight of 0.0001 improves classification performance compared with λc=0\lambda_c=0. Further, From the results, when $\lambda_c $ reaches 0.01, there is a noticeable improvement in performance.

Table 4. Performance comparison with various λc\lambda_c

Dataset00.00010.0010.010.050.1
GRAZ0267.7768.0468.4469.9369.5969.26
Caltech867785.6586.2986.1286.4186.1286.35

Response to W5: Robustness under Extreme Noisy Correspondence

To verify the robustness of HGR under extreme noisy correspondence, we conduct experiments on five datasets with different numbers of views, where more than half of the views are corrupted during testing. The experimental results demonstrate that HGR consistently achieves the best classification performance across all five datasets, outperforming existing methods even under extreme noisy correspondence. These findings further validate the robustness and effectiveness of HGR in challenging scenarios with severe view corruption.

Table 5. Performance comparison under extreme NC, where NC rate is set to 0.5.

MethodGRAZ02Caltech8677flower17LandUseNUSWIDEOBJ
TMC56.4259.9147.7933.8131.35
UIMC56.7559.6159.6133.5732.90
ECML53.0459.3344.1133.0928.53
CCML55.0659.5054.0429.2829.56
MAMC56.4142.1047.7942.3834.21
ETF60.8154.3769.8537.1436.73
TMCEK52.0258.4638.6032.8528.60
FUML45.9451.155.8848.8038.51
RML55.4055.0666.1748.0933.46
BML64.8658.7080.5155.0040.83
HGR65.5460.2081.2555.9543.28

I appreciate the authors' detailed response. The additional experimental results have addressed my concerns, and I support the acceptance of this paper.

Thank you for your recognition and valuable feedback, which are of great significance to improving our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.