acceptodds
Under review as a conference paper at ICLR 2027

Hub for Outliers, Spokes for Inliers: Uniform Latent Space Construction for Dual-Mismatched Semi-Supervised Learning

Abstract

Semi-supervised learning typically assumes that labeled and unlabeled data share an identical class distribution and label space. However, this setting is often violated: unlabeled data may be imbalanced and contain unknown class samples, causing mismatches in both class distribution and label space. Such dual mismatch leads to majority classes dominating the latent space and unknown class samples being overconfidently misclassified, degrading feature discriminability and pseudo-label quality. To address this, we propose a hub-spoke latent geometry, where known classes are uniformly distributed around a central hub and each class forms compact clusters around its prototype, while the hub serves as a low-evidence region specifically designed for high-uncertainty unknown class samples. Integrated with an evidence-based classifier, this geometry ultimately enhances feature discriminability and uncertainty separation by mitigating majority-class domination through structured feature organization and guiding high-uncertainty unknown class samples toward the low-evidence hub. Extensive experiments show that our method outperforms state-of-the-art methods, with a maximum improvement of **3.25%** across various settings. **Code is available in the supplementary material.**

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Dual-Mismatched Semi-Supervised Learning, Uniform Latent Space Construction, OOD, SSL

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 2 reviews, above 27% of the others.
Reviewer_ee6dOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses dual-mismatched semi-supervised learning, where unlabeled data exhibit both class imbalance and the presence of unknown classes. The authors propose a hub-spoke latent geometry that encourages uniform separation of known-class prototypes and compact within-class representations. The framework also introduces class-adaptive reweighting to improve minority-class learning and an evidential classifier to identify high-uncertainty samples, which are subsequently guided toward a central hub. The paper provides a theoretical analysis of the low-evidence region around the hub and evaluates the method on five image classification benchmarks. The results show consistent improvements in average classification accuracy over existing SSL, long-tailed SSL, and open-set SSL methods, although the magnitude of improvement varies across datasets and settings.

Critical strengths
  1. The paper addresses an important and practical problem in semi-supervised learning. Considering class distribution mismatch and unknown-class contamination simultaneously is well motivated, as both can arise in real-world unlabeled datasets.

  2. The proposed hub-spoke geometry provides an intuitive way to organize known-class representations while separating potentially unknown samples.

  3. The experimental evaluation covers five datasets and includes comparisons with a broad range of SSL, long-tailed SSL, and open-set SSL methods. The authors also provide ablation studies and analyses of minority-class accuracy and pseudo-label quality.

  4. The proposed method achieves the highest average classification accuracy across the evaluated settings, with meaningful improvements in some challenging low-label scenarios.

Critical weaknesses
  1. The methodological novelty is somewhat limited. The proposed framework combines several established techniques, including prototype-based representation regularization, class-adaptive reweighting, and evidential uncertainty estimation. The overall approach appears somewhat ad hoc, and the paper does not sufficiently establish a unified principle explaining why these particular components should work together to address dual mismatch. The ablation studies demonstrate incremental improvements but do not establish whether the proposed integration offers advantages over simpler combinations of existing methods.

  2. The design of the hub-spoke geometry lacks sufficient justification. In particular, it is unclear why samples from potentially diverse unknown classes should be encouraged to cluster around a single central hub. The theoretical analysis establishes a low-evidence region under certain assumptions, but does not explain why this geometry is optimal or particularly effective for handling unknown classes.

  3. The unknown-class identification strategy may conflict with the objective of improving minority-class representations. The method selects the most uncertain unlabeled samples as unknown-class candidates. However, minority-class samples may also exhibit high uncertainty due to limited representation during training. Incorrectly pulling these samples toward the hub could reinforce class imbalance. The paper does not sufficiently examine this potential failure mode.

Questions for authors

How does the method distinguish unknown-class samples from difficult known-class samples, especially those belonging to minority classes? Can the authors report unknown-class detection precision and recall, along with false rejection rates for majority and minority classes?

Can the authors provide stronger comparisons with closely related prototype-based SSL methods, including the simplex-based structural approach cited in the paper?

5 comments

We appreciate the reviewer's insightful comments and respond to each point below.

Limited novelty (W1)

Our contribution lies in the unified hub-spoke geometry, which jointly organizes known and unknown classes in a shared latent space. We first enforce uniformly spread spokes to maximize inter-class discriminability. Then, exploiting the hub's separation from the spokes, we construct a low-evidence hub region that further separates known from unknown samples. Unlike methods that separately apply prototype-based learning to known classes and uncertainty filtering to unknown classes, our approach couples both through a single geometry.

Why cluster diverse unknown classes to a single hub (W2)

Our goal is to separate unknown classes from known classes, rather than distinguish unknown classes from each other. From this perspective, pulling high-uncertainty unknown samples toward the hub is a natural design. The hub, defined as the mean of all known-class prototypes, is not associated with any known class and lies away from the known-class prototypes. By encouraging unknown samples to cluster around the hub, our method simultaneously separates known and unknown samples in two complementary spaces: in latent space, unknown samples are drawn toward the hub, away from the known-class spokes; in uncertainty space, the hub-centered low-evidence region distinguishes unknown samples from the high-evidence known samples. This allows diverse unknown classes to share a common representation region without requiring discrimination among them. As shown in Fig. 2d, guiding identified unknown samples toward the hub yields a clearer separation between known and unknown uncertainty distributions, validating this desgin.

Conflict with minority-class learning (W3)

To mitigate this risk, we set kk to a small value (10% of the batch) and select only the high-uncertainty unknown-class samples, rather than using uncertainty to filter out all unknown-class samples. This selection reduces the risk of pulling uncertain minority-class samples toward the hub and reinforcing class imbalance. We further mitigate the under-representation of minority classes through (i) the hub-spoke geometry, which promotes high inter-class separation and intra-class compactness, and (ii) pre-training and class-adaptive reweighting (Eq. 9), which amplify gradients for minority classes. These measures increase minority-class accuracy from 33% to 64% (Fig. 4a, lb4), while maintaining a low known-class false-rejection rate of 0.028 on CIFAR-10 (lb25).

Unknown-class detection metrics and false-rejection rates (Q1)

On CIFAR-10 (lb25), unknown-class detection achieves precision 0.78590.7859, recall 0.15030.1503, and F1 0.25230.2523, with an overall known-class false-rejection rate of 0.0280.028 (majority: 0.02790.0279; minority: 0.02800.0280). The high precision (0.78590.7859) and the low false-rejection rate (0.0280.028) confirms that we accurately target high-uncertainty unknown-class samples without sacrificing known-class samples. However, this high precision comes at the cost of recall, as our small kk selection strategy focuses on the most uncertain candidates to reduce the risk of misidentifying difficult known-class samples as unknown. By encouraging selected unknown samples to approach the low-evidence hub and suppressing their evidence through Eq. (10), the evidential classifier learns to assign low evidence to other unknown samples as well, even when they are not explicitly selected.

Comparison with prototype-based methods (Q2)

Existing methods use prototypes or geometric anchors for different purposes. Prototypical Contrastive Learning [1] uses cluster-level prototypes and ProtoNCE to learn semantic representations. TRAILER [2] aligns representations with pre-assigned ETF targets, while SAGE [3] uses fixed simplex ETF vectors as a class-agnostic coordinate frame for graph-based structural inference.

Our hub differs: as the mean of all known-class prototypes, it is an information-free, low-evidence anchor far from the spokes, giving unknown classes a dedicated region separated from known classes in two coupled senses. In label space, uniformly spread spokes maximize inter-class separation while the hub stays far from every spoke, so known and unknown decision boundaries never overlap. In uncertainty, the low-evidence hub keeps high-uncertainty unknown samples pulled toward it uncertain without degrading known-class evidence. This dual separation, combining known-class spokes with an unknown-class hub, unifies class-imbalance handling and unknown-class detection in one geometry, which none of the prototype-based methods above achieve.

[1] ICLR’21-Prototypical Contrastive Learning of Unsupervised Representations

[2] CVPR’24-Targeted Representation Alignment for Open-World Semi-Supervised Learning

[3] ICML’26-Beyond Distribution Estimation: Simplex Anchored Structural Inference Towards Universal Semi-Supervised Learning

Dear Reviewer ee6d,

Thanks for your time and effort in reviewing this paper. Your comments are valuable to improve the quality of this paper. As the author-reviewer discussion period ends in 48 hours, we kindly request that you review the rebuttal and confirm whether your concerns have been addressed. Thanks a lot.

Regards from the authors.

Dear Reviewer ee6d,

Thanks for your time and effort in reviewing this paper. Your comments are valuable to improve the quality of this paper. As the author-reviewer discussion period ends in 24 hours, we kindly request that you review the rebuttal and confirm whether your concerns have been addressed. Thanks a lot.

Regards from the authors.

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

This paper studies dual-mismatched semi-supervised learning (DMSSL), where unlabeled data exhibit both class imbalance and label-space mismatch, including samples from unknown classes. The proposed method organizes the latent space into a hub-and-spoke geometry: known classes occupy separated spokes, while unknown samples are encouraged to concentrate around a central hub. Global regularization separates labeled class prototypes angularly and encourages them to lie at equal distances from the hub. Local regularization makes samples compact and radially uniform around their class prototypes. A class-adaptive weighting mechanism uses the displacement between labeled-only and overall class prototypes to reweight unlabeled samples from minority classes. For unknown-class identification, an evidential classifier models predictions with a Dirichlet distribution and computes uncertainty as (u=C/S). The 10% most uncertain samples in each batch are treated as unknown-class candidates and pulled toward the hub. The theoretical analysis argues that, under a Lipschitz continuity assumption, this encourages a region of low evidence around the center.

Critical strengths
  1. The setting is practically relevant. Unlabeled data can be both imbalanced and contaminated by unknown classes. Studying these two issues together is useful, and the experiments include 13 baselines spanning conventional SSL, long-tailed SSL, and open-set SSL.
  2. Separating known classes around a shared center while directing unknown samples toward that center gives the method an intuitive structure. Using the mean of labeled class prototypes as the hub also connects naturally to the intended low-evidence region.
Critical weaknesses
  1. Selecting a fixed 10% of each batch as unknown candidates is a restrictive assumption. The unknown-class proportion is generally unavailable and may differ substantially from 10%. More concerning, known minority-class samples are often highly uncertain during early and intermediate training.
  2. The theoretical result is limited, and its presentation needs correction. Lemma 1 and Theorem 1 largely follow from the definition of Dirichlet uncertainty and a Lipschitz bound. Since Loc\sqrt{L_{oc}} is an average squared distance, at least one selected sample lies within Loc\sqrt{L_{oc}} of the hub; applying the Lipschitz assumption then gives the stated result. The bound also depends on an unconstrained Lipschitz constant LocL_{oc} making its practical implications difficult to assess.
  3. The training procedure is complex. It involves three stages, including 60 epochs of the first warmup and 20 epochs of the second, two sets of prototypes, and two classification heads. A clearer account of which components are essential, and how much tuning they require, would improve the practical case for the method.
Questions for authors

Why do the SSL methods perform so much worse than supervised learning on Food-101 and ImageNet-100? The proposed method obtains 19.82% and 27.52%, respectively, compared with supervised results of 42.56% and 46.61%. What accounts for these gaps, and for the improvement of only 0.08 percentage points over FixMatch on ImageNet-100 with 10 labels?

3 comments

We appreciate the reviewer's insightful comments and respond to each point below.

Fixed top-10% unknown-candidate selection (W1)

The 10% selection ratio is a cautious candidate-selection strategy, not an assumption that unknown classes account for 10% of each batch. Since the unknown-class proportion is unavailable and minority-class samples may also exhibit high uncertainty due to limited representations, we use a small kk to focus on the most uncertain candidates and reduce the risk of mistakenly treating known-class samples as unknown. We then encourage these selected samples to approach the low-evidence hub and train the evidential classifier to suppress their evidence, which enables generalization to other unknown samples without requiring all of them to be explicitly selected. This strategy maintains a low known-class false-rejection rate of 0.028 on CIFAR-10 (lb25). Meanwhile, Table 2 shows that UCIM brings an average improvement of 2.06 percentage points on CIFAR-10, demonstrating the effectiveness of this selective unknown-class modeling.

Limited theoretical result (W2)

We agree Lemma 1 and Theorem 1 are not intended as a tight or practically calibrated bound. They give a simple sufficient characterization of the mechanism. Lemma 1 bounds the evidence of top-kk candidates by εO,k=C(1/uk−1)\varepsilon_{O,k}=C(1/u_k-1), and Theorem 1 combines this with the hub-clustering term LocL_{oc} and Lipschitz continuity to show sup⁡z∈H(r)gc(z)≤εO,k+Lc(r+Loc)\sup_{z\in\mathcal{H}(r)}g_c(z)\le \varepsilon_{O,k}+L_c\big(r+\sqrt{L_{oc}}\big). Our claim is not a sharp numerical guarantee, but that this formally connects the two algorithmic ingredients, top-kk uncertainty selection and hub-centered clustering, to the intended low-evidence region around the hub.

Complex training procedure (W3)

The progressive warm-up schedule and the two complementary prediction heads are core components of our method. Specifically, the first warm-up (60 epochs) establishes initial feature representations and global geometry through contrastive learning and geometry regularization, while the second warm-up (20 epochs) introduces classification objectives to establish basic classification capability and stabilize pseudo-labels and uncertainty estimates. This progressive design provides reliable representations and predictions before activating the full objective. The classification head predicts known-class probabilities and generates pseudo-labels, whereas the evidence head estimates uncertainty to identify unknown-class samples. These two heads are essential for jointly addressing class distribution and label space mismatches. Furthermore, labeled-class prototypes provide reliable references for global geometry construction, whereas overall class prototypes serve as references for local geometry regularization.

Regarding tuning, we use the same 60/20-epoch schedule and λ=0.1\lambda=0.1 across all datasets and settings. The sensitivity analysis in Appendix D.1 further demonstrates robustness to kk and vv. Specifically, on CIFAR-10, the gains of k=10%k=10\% and k=20%k=20\% over k=0%k=0\% differ by only 1.1 percentage points on average across the three labeled settings, while varying vv results in an average standard deviation of only 0.99%. We will clarify these design choices and their respective roles in Sec 4.3.

Why SSL methods underperform supervised learning (Q1)

The performance gap mainly reflects the intrinsic difficulty of learning on more complex datasets with very limited labeled data. Food-101 contains 101 fine-grained classes, while ImageNet-100 exhibits substantial visual diversity, making reliable feature learning and pseudo-label generation more challenging. With only 10 or 25 labeled samples per class, the limited supervision further constrains representation learning, while unknown-class samples introduce additional noise into the unlabeled data. These factors make semi-supervised learning more challenging than supervised learning on these datasets. On CIFAR-100, our method improves performance by 3.25 percentage points over FixMatch. In comparison, the improvement on ImageNet-100 is only 0.08 percentage points (27.52±0.3027.52\pm0.30 vs. 27.44±0.8127.44\pm0.81), suggesting that the benefit is limited under the greater complexity of this dataset and scarce supervision.

Dear Reviewer SNmR,

Thanks for your time and effort in reviewing this paper. Your comments are valuable to improve the quality of this paper. As the author-reviewer discussion period ends in 48 hours, we kindly request that you review the rebuttal and confirm whether your concerns have been addressed. Thanks a lot.

Regards from the authors.

Dear Reviewer SNmR,

Thanks for your time and effort in reviewing this paper. Your comments are valuable to improve the quality of this paper. As the author-reviewer discussion period ends in 24 hours, we kindly request that you review the rebuttal and confirm whether your concerns have been addressed. Thanks a lot.

Regards from the authors.

Comments on the paper

6 comments

We appreciate the reviewer's insightful comments and respond to each point below.

Interpretation of Eqs. (7) and (9) (W1/Q1)

Regarding Eq. (7), we would like to clarify that our intention is not to establish a strict guarantee of directional uniformity through this term alone. Rather, Eq. (7) penalizes the overall directional alignment within each class. In particular, when within-class feature directions exhibit a strong positive alignment, minimizing this term encourages their average pairwise cosine similarity to decrease, thereby alleviating excessive directional concentration. In practice, this regularization is jointly optimized with the intra-local loss, which encourages compact and discriminative feature representations around the corresponding class prototypes. We therefore view Eq. (7) as an auxiliary directional regularizer that contributes to a more structured within-class feature distribution, rather than a standalone objective guaranteeing uniform angular coverage.

Regarding Eq. (9), the cosine similarity is direction-based rather than magnitude-based, so the larger deviation distances in Fig. 3 do not by themselves imply larger weights. Rather, the weight is large when a pseudo-labeled sample's displacement qiq_i is directionally aligned with the class-level deviation Δyi\Delta_{y_i}. The observation is that minority classes exhibit systematically larger deviation (Fig. 3), and their samples are displaced in the direction from the labeled prototype toward the global prototype, so qiq_i and Δyi\Delta_{y_i} are more consistently aligned and the cosine, hence the weight, is larger.

Isolating the contribution of hub attraction. (W2/Q2)

To isolate the contribution of hub attraction, we conduct an ablation that retains the original uncertainty-based unknown-class candidate selection and the entire training schedule, but removes the attraction of selected unknown-class candidates toward the hub. The results are summarized below.

Labeled samples/classCIFAR-10 (w/o Hub)CIFAR-10 (Full Model)GainCIFAR-100 (w/o Hub)CIFAR-100 (Full Model)Gain
454.1756.89+2.7229.7030.43+0.73
1068.1870.96+2.7841.5444.37+2.83
2577.4278.05+0.6353.2653.53+0.27

We further evaluate unknown-class detection on CIFAR-10 with 25 labeled samples per class:

MetricCIFAR-10 (w/o Hub)CIFAR-10 (Full Model)
Precision0.68640.7859
Recall0.12270.1503
F10.21530.2523
Known-class false rejection0.03990.0280

These results demonstrate that hub attraction provides consistent improvements in known-class classification and unknown-class detection, even when uncertainty-based candidate selection remains unchanged. In particular, it improves unknown-class detection precision by 9.95 percentage points while reducing the known-class false-rejection rate by 1.19 percentage points. This supports the complementary role of hub attraction in improving known–unknown discrimination beyond candidate selection alone.

Aspects Acknowledged by the Reviewers

  1. Practical relevance of the problem. Reviewers recognized the importance of DMSSL, which jointly considers class distribution mismatch and unknown-class contamination in unlabeled data.

  2. Intuitive methodological design. The hub-spoke geometry was recognized as an intuitive structure for organizing known classes around separated spokes while providing a central anchor for modeling unknown classes.

  3. Broad empirical evaluation. Reviewers acknowledged the evaluation on five datasets against 13 baselines covering conventional SSL, long-tailed SSL, and open-set SSL, supported by ablation studies and sensitivity analyses.

  4. Empirical effectiveness. Reviewers highlighted the consistent performance improvements, including the maximum gain of 3.25 percentage points, the minority-class accuracy improvement from 33% to 64%, and the positive contribution of UCIM in the progressive ablation study.

Concerns Raised and Our Responses

  1. Methodological novelty and the rationale for hub-centered unknown-class modeling. Reviewers questioned how the components form a unified framework and why diverse unknown classes should cluster around one hub. We clarified that our goal is to separate unknown classes from known classes, rather than distinguish unknown classes from each other. The hub-spoke geometry achieves this separation in both latent and uncertainty spaces. We also expanded comparisons with Prototypical Contrastive Learning, TRAILER, and SAGE to clarify the differences in their uses of prototypes and geometric anchors.

  2. Potential negative effects on minority classes. Reviewers questioned the fixed 10% candidate-selection ratio and the risk of treating uncertain minority-class samples as unknown. We clarified that this ratio is a cautious selection strategy, not an assumption about the unknown-class proportion. We further reported a low overall known-class false-rejection rate of 0.028 and showed that minority-class accuracy improves from 33% to 64%.

  3. Unknown-class detection metrics. In response to requests for detection metrics, we reported precision of 0.7859, recall of 0.1503, and F1 of 0.2523, with an overall known-class false-rejection rate of 0.028. We explained that the relatively low recall results from the cautious small-kk strategy, while hub guidance and evidential learning allow the classifier to generalize to unselected unknown samples.

  4. Independent contribution of hub attraction. To directly isolate its effect, we conducted a controlled ablation that retains the same candidate selection and training schedule but removes hub attraction. Hub attraction consistently improves classification accuracy across CIFAR-10 and CIFAR-100 settings, increases unknown-class detection precision from 0.6864 to 0.7859, and reduces the known-class false-rejection rate from 0.0399 to 0.028.

  5. Theoretical justification and interpretation of geometric objectives. We clarified that the theoretical result provides a sufficient characterization of the low-evidence region rather than a tight numerical guarantee. We also refined the interpretations of Eqs. (7) and (9): Eq. (7) acts as an auxiliary directional regularizer rather than guaranteeing uniform angular coverage, while Eq. (9) emphasizes directional alignment when reweighting minority-class samples.

  6. Training complexity and hyperparameter sensitivity. We explained the respective roles of the two warm-up stages, classification and evidence heads, and the two types of prototypes. We also clarified that the same 60/20-epoch schedule and λ=0.1\lambda=0.1 are used across datasets and settings, with sensitivity analyses supporting robustness to key hyperparameters.

  7. Performance gaps on complex datasets. We attributed the larger gap between SSL and supervised learning on Food-101 and ImageNet-100 to their greater intrinsic complexity and limited labeled data.

Dear Reviewers,

Thanks for your time and effort in reviewing this paper. Your comments are valuable to improve the quality of this paper. As the author-reviewer discussion period will end in about 3 days, we gently request that you read the rebuttal and check whether your concerns are addressed. Thanks again for your time and help in improving the quality of this paper.

Regards from the authors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.