Pith. sign in

REVIEW 4 major objections 5 minor 56 references

Diversify and Conquer: Open-set Disagreement for Robust Semi-supervised Learning with Outliers

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read When labels are scarce, disagreement among models finds the outliers

desk verdict DAC is a solid, well-engineered OSSL method that deserves review, but the central open-set-disagreement hypothesis is under-tested and the abstract overstates the CIFAR-10 results. read the letter →

arxiv 2505.24443 v1 pith:V5C2Z4PS submitted 2025-05-30 cs.CV cs.LG

Heejo Kong , Sung-Jin Kim , Gunho Jung , Seong-Whan Lee This is my paper
classification cs.CVcs.LG
keywords open-setsemi-supervisedlearningoutlierdetectionpredictiondisagreementmulti-headensemblemutualinformationconsensusscoresparselabelsFixMatch
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Open-set semi-supervised learning (OSSL) asks a model trained on a few labeled classes to cope with unlabeled data that also contains unknown classes, or outliers. Existing OSSL methods judge a sample by how uncertain a single model is about it, but that test collapses when there are very few labels, because the model is then uncertain about almost everything and rejects useful inliers. This paper argues that a more reliable signal is disagreement: train several heads that are deliberately biased in different ways toward the unlabeled data; inliers keep the heads in agreement while outliers split them. The authors build this signal cheaply, with one shared feature extractor and a set of heads, and use consensus scores to softly down-weight suspected outliers in the standard SSL objective. The reported result is state-of-the-art OSSL accuracy on CIFAR, ImageNet, and Semi-iNat benchmarks, with the largest gains precisely in the sparse-label regime where single-model methods fail.

What carries the argument

The central object is the multi-head predictor $\tilde{g} = \{g_1, \ldots, g_K\}$ on a shared encoder, trained by the diversification objective $L_{\text{div}} = \sum_i L_{\text{ssl}}(g_i) + \lambda_{\text{mi}} \sum_{i \neq j} L_{\text{mi}}(g_i, g_j)$, where $L_{\text{mi}}$ is the mutual information between head predictions on unlabeled embeddings. The consensus score $\hat{s}_b = \mathbb{E}_{(g_i,g_j)}[\exp(-\|p(g_i|z_b) - p(g_j|z_b)\|)]$ quantifies open-set disagreement: values near 1 mean the heads agree (inlier), near 0 mean they disagree (outlier). The scores feed three downstream mechanisms: a min-max normalization plus EMA memory queue for stability, a soft-rejection weight $\omega(u_b) = (s_b/\tau_{\text{open}})^{t_\omega}$ below an Otsu-derived threshold $\tau_{\text{open}}$, and an open-set knowledge distillation loss $L_{\text{kd}}$ that pushes the shared encoder to separate low-consensus samples from known-class clusters in embedding space.

What would settle it

Train DAC on a benchmark with 1–2 labeled samples per class and plot the consensus-score histograms for inliers and outliers separately; if the two distributions overlap so much that an outlier-detection AUROC near chance, the open-set disagreement signal has dissolved and the robustness claim for underspecified labels is refuted. A cheaper check is to ablate the mutual information loss on CIFAR-10 with 5 labels per class and measure whether inlier disagreement grows as much as outlier disagreement, as the paper's limitation note predicts.

Watch

Extended reading notes

Core claim

The paper's central claim is that open-set disagreement—prediction inconsistency across models that are each biased differently toward the unlabeled distribution—is a usable detector for unknown-class samples, and that it remains usable when labeled data is underspecified. The authors show experimentally that SSL models trained on the same labeled set but different unlabeled mixtures produce similar predictions for inliers but diverging predictions for outliers, and they elevate this observation into a training objective. Their framework, DAC, creates K divergent heads on a shared encoder, drives the heads to be statistically independent on unlabeled data through a mutual information loss, and defines a consensus score per sample as the average exponential-negative-L1 distance between all head pairs. Low consensus marks an outlier; the scores are then distilled into the shared feature extractor via a graph-based knowledge distillation loss and used to softly reweight the unsupervised loss so that likely outliers contribute little. The paper reports that this outperforms existing open-set SSL methods across CIFAR-10/100, ImageNet-30, and Semi-iNat-2021, and that unlike single-model detectors it does not over-reject inliers when labels are very scarce.

Load-bearing premise

The load-bearing premise is that multiple SSL models trained with the same labels will keep agreeing on inliers while disagreeing on outliers; if labels are so scarce that inliers also split the heads, the consensus score can no longer separate known from unknown samples.

Editorial extensions

If this is right

  • OSSL can be solved without a detector trained on labeled data alone; the consensus of diverse heads trained on unlabeled data supplies the outlier signal, so the method does not collapse when labels are too sparse to train a reliable single model.
  • The multi-head design delivers ensemble-like disagreement at roughly single-model inference cost, making disagreement-based outlier detection practical within a single training run.
  • Soft rejection—down-weighting rather than discarding likely outliers—preserves ambiguous inliers and yields better closed-set accuracy than hard rejection, directly countering the over-rejection failure of prior methods.
  • DAC improves closed-set accuracy when attached to several SSL baselines (MixMatch, SimMatch, SoftMatch), indicating the disagreement mechanism is a general add-on rather than a FixMatch-specific fix.
  • The advantage over single-model OSSL methods widens as labeled samples become fewer and as the outlier fraction grows, which is the regime where prior detect-and-filter methods break down.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If open-set disagreement is as fundamental as the paper claims, the same consensus-score computation could be applied at inference time to flag novel classes in an open-set recognition setting, not just during SSL training.
  • The uniform mutual information penalty is a blunt instrument: a variant that applies diversity only to low-confidence samples could avoid the 'excessive separation' the authors themselves report as a limitation and likely improve the CIFAR-10 results.
  • The pairwise L1-distance consensus could be replaced by a learned agreement metric or a softmax over agreement values; which aggregation is best is an empirical question the paper leaves open.
  • The Otsu threshold presumes a bimodal consensus-score histogram; adapting the threshold to the expected outlier fraction (e.g., when outliers are rare) would test whether the method degrades gracefully outside the balanced setting.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. This manuscript proposes DAC, an open-set semi-supervised learning method that trains a shared feature extractor with multiple classification heads. Each head is trained with a standard SSL objective plus a pairwise mutual-information penalty that encourages divergent predictions on unlabeled data; the average pairwise L1 agreement among heads (Eq. 7) defines a consensus score used to identify outliers, reweight the unsupervised loss (Eq. 13), and distill a virtual unknown class into the feature extractor (Eqs. 10-12). The paper reports experiments on CIFAR-10/100, ImageNet-30, and Semi-iNat-2021, comparing with standard SSL and OSSL baselines, and includes ablations, hyperparameter sensitivity, mismatch-ratio and correlated-outlier experiments.

Significance. If the central hypothesis holds—that differently biased heads remain consistent on inliers but disagree on outliers even with very few labels—DAC would be a practically useful single-training alternative to ensembles for OSSL. Strengths of the submission include released code, a unified USB-based evaluation test bed, extensive ablations, sensitivity analyses, and experiments on a real-world dataset (Semi-iNat-2021), as well as a clearly described soft-rejection mechanism. However, the paper's headline claim of significant state-of-the-art improvement is not uniformly supported by its own tables (CIFAR-10 open-set accuracy is lower than IOMatch), and the core consensus-score mechanism lacks a direct quantitative evaluation of inlier/outlier separation. The method's novelty depends on whether inlier consistency survives the mutual-information penalty in the low-label regime, which the current evidence does not yet establish.

major comments (4)
  1. [III-B, Eq. (6)] The central outlier signal is the consensus score of Eqs. (7)-(8), whose validity requires that the K heads agree on inliers and disagree on outliers (Section II-B). However, Lmi in Eq. (6) is minimized over all pairs of heads on all unlabeled data, including inliers; the only force keeping inlier predictions consistent is the supervised loss on 5-25 labeled samples per class. The Limitations section concedes that mutual information "encourages diverse predictions indiscriminately for both inliers and outliers, potentially causing excessive separation." Figures 6-7 are qualitative heatmaps and histograms; they do not quantify how well the consensus score separates inliers from outliers. Please report a threshold-free separation metric (AUROC or PR-AUC) of s_b against ground-truth inlier/outlier labels across the label-count settings, or the mean inlier-vs-outlier consensus gap with error bars. Without this, the reweighting in Eq. (13) might be suppressing inliers precisely in the underspecified regime that the paper targets.
  2. [IV-A, Evaluation metrics] The manuscript states that "we report the best results across all training epochs" and that each method is evaluated using "its best checkpoint model based on closed-set performance." This is test-set-based model selection: the closed-set accuracy used to pick the checkpoint is computed on known-class test data. This can inflate all numbers and, more importantly, can bias the open-set balanced accuracy ranking, because the chosen checkpoint is not selected on outlier-detection performance. Please either report results at the final epoch, use a held-out validation split for model selection, or otherwise justify why this protocol does not advantage the proposed method over baselines.
  3. [Abstract; IV-B, Tables 1-2] The abstract claims the method "significantly surpasses state-of-the-art OSSL methods across various protocols," but Table 2 shows the opposite on CIFAR-10: IOMatch open-set balanced accuracy is 74.8/75.1/79.0 for 5/10/25 labels per class whereas DAC is 70.1/71.5/75.1, so DAC is lower in all three columns. The text's explanation that all hyperparameters were tuned on CIFAR-50-500 is plausible but not demonstrated. Please either add per-dataset tuned results for CIFAR-10, report the fixed-hyperparameter results as a separate robustness check, or qualify the abstract so it does not claim uniform state-of-the-art improvement.
  4. [II-B vs III-B, Fig. 2 vs Fig. 6] The motivating experiment (Fig. 2) trains fully independent SSL models on disjoint unlabeled subsets, whereas the implemented DAC uses a shared encoder and all heads see the same unlabeled batches, with diversity induced only by Lmi and the gradient-blocking scheme. The authors note in Section IV-C that their heads show "more similar predictions overall" than the independent models in Fig. 2. Because the central claim is that open-set disagreement emerges in the implemented architecture, the paper should provide a quantitative comparison of the consensus distributions (e.g., overlap of inlier/outlier score histograms or AUROC) between the independent-model setup and the shared-encoder DAC setup, rather than relying on the qualitative Fig. 6 heatmaps.
minor comments (5)
  1. [Eq. (6)] There is a typo in Eq. (6): the second factor in the product of marginals should be p(gj|zb), not p(gi|zb).
  2. [Throughout] Please fix typographical errors: "corupt" in Section I, "datsasets" in Section IV-A, "I NTRODUCTION" in the header, and "Mismatched" in Section IV-E.
  3. [Fig. 2 caption] The Fig. 2 caption says "CIFAR-50-500" while Section IV uses "CIFAR-50" with various labeled-sample counts; unify the notation (e.g., CIFAR-50-500 vs CIFAR-50-100) to avoid confusion.
  4. [Algorithm 1 and Eq. (14)] In Algorithm 1, Ltotal is first set to Ls + Ldiv and then augmented with Lu and Lkd; for clarity, align the pseudocode with Eq. (14) by defining the complete total loss once and stating which terms are active during the warm-up phase.
  5. [Related works] The method's name overlaps with the title of reference [10] ("Diversify & Conquer"); a sentence distinguishing the two uses would help avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; DAC's outlier signal is an emergent self-training quantity validated on held-out ground truth, not a fitted parameter or self-citation chain.

full rationale

The paper's consensus scores are computed from the trained divergent heads and then used to reweight the unsupervised loss and to distill open-set targets into the shared feature extractor. This is a self-referential training loop, but it is standard self-training rather than circular reasoning: the final claims are evaluated on held-out open-set test data via balanced accuracy (Eq. 15), and the consensus scores are never fitted to ground-truth outlier labels. The hyperparameters are fixed before evaluation, and the ablation studies show the incremental contribution of each loss component. The only self-citation, reference [27], is used as a baseline and as a source of experimental protocols (e.g., correlated-outlier settings), not as a justification of the method's correctness. The authors' own limitation that the mutual-information loss indiscriminately separates inliers and outliers is an acknowledged weakness that could undermine the mechanism, but it does not make the derivation circular; it is an empirical risk. No equation reduces by construction to its own inputs, and no load-bearing claim is imported solely from the authors' prior work.

Assumptions & free parameters 9 free parameters · 4 assumptions · 1 invented entities

The method rests on several hand-picked hyperparameters (tuned on one dataset), the unproven open-set disagreement hypothesis, and a virtual unknown class used to shape the feature space. These are typical for SSL methods but should be stated explicitly.

free parameters (9)
  • lambda_u (supervised loss weight) = 1.0
    Set to a standard value; not ablated.
  • lambda_mi (mutual information weight) = 0.5
    Tuned on CIFAR-50-500; sensitivity analysis shows performance is sensitive to this value.
  • lambda_kd (distillation weight) = 1.5
    Tuned on CIFAR-50-500; values above 1 are robust.
  • tau (pseudo-label confidence threshold) = 0.95
    Standard FixMatch threshold, transferred without tuning.
  • alpha (EMA momentum for consensus scores) = 0.9
    Chosen as a default; not ablated.
  • t_e (sharpening temperature for memory bank affinities) = 0.1
    Fixed hyperparameter.
  • t_w (soft rejection temperature) = 1.5
    Tuned on CIFAR-50-500; Figure 9 shows sensitivity, with peak around this value.
  • K (number of divergent heads) = 10
    Sensitivity analysis in Figure 8 shows performance plateaus around K=10.
  • M (memory bank size) = 256 * mu * B
    Set proportional to batch size; not ablated.
assumptions (4)
  • domain assumption Multiple SSL models or heads trained on unlabeled data agree on inliers and disagree on outliers (open-set disagreement hypothesis).
    Stated in Section II-B, used to justify consensus score Eq. (7); only empirical motivation is provided, no proof.
  • standard math The batch-based empirical KL divergence (Eq. 6) is a reliable estimator of mutual information between heads.
    Borrowed from prior work [10,32,33]; its adequacy is not analyzed in this paper.
  • domain assumption Otsu's threshold on the consensus score distribution separates inliers from outliers.
    Used in Section III-C, Eq. (13); assumes a bimodal score distribution.
  • domain assumption FixMatch's assumptions (confident pseudo-labels are correct, strong augmentations preserve semantics) hold in the OSSL setting.
    The SSL backbone relies on FixMatch [42]; if pseudo-labels are noisy for inliers, the whole framework degrades.
invented entities (1)
  • Virtual (C+1)-th unknown class
    purpose: Represents all outlier samples in the open-set target distribution q_b (Eq. 10) for knowledge distillation.
    A modeling construct that aggregates all unknown classes into one generic category; there is no external falsifiable prediction tied to it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Diversify and Conquer: Open-set Disagreement for Robust Semi-supervised Learning with Outliers." pith.science (2026). https://pith.science/paper/V5C2Z4PS

@misc{pith2026250524443,
  author       = {Pith},
  title        = {Pith review of: Diversify and Conquer: Open-set Disagreement for Robust Semi-supervised Learning with Outliers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V5C2Z4PS}},
  note         = {Machine review of arXiv:2505.24443}
}
read the original abstract

Conventional semi-supervised learning (SSL) ideally assumes that labeled and unlabeled data share an identical class distribution, however in practice, this assumption is easily violated, as unlabeled data often includes unknown class data, i.e., outliers. The outliers are treated as noise, considerably degrading the performance of SSL models. To address this drawback, we propose a novel framework, Diversify and Conquer (DAC), to enhance SSL robustness in the context of open-set semi-supervised learning. In particular, we note that existing open-set SSL methods rely on prediction discrepancies between inliers and outliers from a single model trained on labeled data. This approach can be easily failed when the labeled data is insufficient, leading to performance degradation that is worse than naive SSL that do not account for outliers. In contrast, our approach exploits prediction disagreements among multiple models that are differently biased towards the unlabeled distribution. By leveraging the discrepancies arising from training on unlabeled data, our method enables robust outlier detection even when the labeled data is underspecified. Our key contribution is constructing a collection of differently biased models through a single training process. By encouraging divergent heads to be differently biased towards outliers while making consistent predictions for inliers, we exploit the disagreement among these heads as a measure to identify unknown concepts. Our code is available at https://github.com/heejokong/DivCon.

Figures

Figures reproduced from arXiv: 2505.24443 by the authors.

Figure 1
Figure 1. Comparison results for various open-set SSL algorithms (colored [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Qualitative results for multiple models trained with standard SSL baseline [42] on the [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. An overview of the proposed framework, DAC, for the unlabeled data training. Our key components, contrasting with conventional SSL, include [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: A conceptual illustration of estimating consensus scores with divergent [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: UMAP visualizations of embeddings obtained from the ablated models. Black points denote outliers, while the other colored points represent distinct [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 7
Figure 7. Figure 7: Histogram plots of the consensus scores (Eq. (8)) acquired from [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 10
Figure 10. Figure 10: Analysis of hyperparameter sensitivity for the weights of objective [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 9
Figure 9. Figure 9: Results of closed-set and open-set accuracy (%) for models trained [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 11
Figure 11. Figure 11: Comparison results for various open-set SSL algorithms trained on [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 54 canonical work pages

  1. [1]

    O’Connor, and Kevin McGuinness

    Eric Arazo, Diego Ortego, Paul Albert, Noel E. O’Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi- supervised learning. In Proc. International Joint Conference on Neural Networks (IJCNN), pages 1–8, 2020

    work page 2020

  2. [2]

    Cubuk, Alex Kurakin, Ki- hyuk Sohn, Han Zhang, and Colin Raffel

    David Berthelot, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Ki- hyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with distribution matching and augmentation anchoring. In Proc. International Conference on Learning Representations (ICLR) , 2020

    work page 2020

  3. [3]

    Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel

    David Berthelot, Nicholas Carlini, Ian J. Goodfellow, Nicolas Papernot, Avital Oliver, and Colin Raffel. Mixmatch: A holistic approach to semi- supervised learning. In Advances in Neural Information Processing Systems (NeurIPS), pages 5050–5060, 2019

    work page 2019

  4. [4]

    Kay Henning Brodersen, Cheng Soon Ong, Klaas Enno Stephan, and Joachim M. Buhmann. The balanced accuracy and its posterior distribu- tion. In Proc. International Conference on Pattern Recognition (ICPR) , pages 3121–3124, 2010

    work page 2010

  5. [5]

    Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning

    Paola Cascante-Bonilla, Fuwen Tan, Yanjun Qi, and Vicente Ordonez. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. In Proc. AAAI Conference on Artificial Intelligence (AAAI) , pages 6912–6920, 2021

    work page 2021

  6. [6]

    Softmatch: Addressing the quantity-quality tradeoff in semi-supervised learning

    Hao Chen, Ran Tao, Yue Fan, Yidong Wang, Jindong Wang, Bernt Schiele, Xing Xie, Bhiksha Raj, and Marios Savvides. Softmatch: Addressing the quantity-quality tradeoff in semi-supervised learning. In Proc. International Conference on Learning Representations (ICLR) , 2023

    work page 2023

  7. [7]

    Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E. Hinton. Big self-supervised models are strong semi- supervised learners. In Advances in Neural Information Processing Systems (NeurIPS), 2020

    work page 2020

  8. [8]

    Exploring simple siamese representation learning

    Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 15750–15758, 2021

    work page 2021

Show all 56 references
  1. [9]

    Semi- supervised learning under class distribution mismatch

    Yanbei Chen, Xiatian Zhu, Wei Li, and Shaogang Gong. Semi- supervised learning under class distribution mismatch. In Proc. AAAI Conference on Artificial Intelligence (AAAI) , pages 3569–3576, 2020

    2020

  2. [10]

    Daesol Cho, Seungjae Lee, and H. Jin Kim. Diversify & conquer: Outcome-directed curriculum RL via out-of-distribution disagreement. In Advances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  3. [11]

    The cityscapes dataset for semantic urban scene under- standing

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene under- standing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CV...

    2016

  4. [12]

    Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V

    Ekin D. Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V . Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3008–3017, 2020

    2020

  5. [13]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei- Fei. Imagenet: A large-scale hierarchical image database. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 248–255, 2009

    2009

  6. [14]

    Semi- supervised learning via weight-aware distillation under class distribution mismatch

    Pan Du, Suyun Zhao, Zisen Sheng, Cuiping Li, and Hong Chen. Semi- supervised learning via weight-aware distillation under class distribution mismatch. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 16410–16420, 2023

    2023

  7. [15]

    Mutexmatch: Semi-supervised learning with mutex- based consistency regularization

    Yue Duan, Zhen Zhao, Lei Qi, Lei Wang, Luping Zhou, Yinghuan Shi, and Yang Gao. Mutexmatch: Semi-supervised learning with mutex- based consistency regularization. IEEE Trans. on Neural Networks and Learning Systems, pages 8441–8455, 2024

    2024

  8. [16]

    Ssb: Simple but strong baseline for boosting performance of open-set semi- supervised learning

    Yue Fan, Anna Kukleva, Dengxin Dai, and Bernt Schiele. Ssb: Simple but strong baseline for boosting performance of open-set semi- supervised learning. In Proc. IEEE International Conference on Com- puter Vision (ICCV) , pages 16068–16078, 2023

    2023

  9. [17]

    Semi-supervised learning by entropy minimization

    Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In Advances in Neural Information Processing Systems (NeurIPS), pages 529–536, 2004

    2004

  10. [18]

    Safe deep semi-supervised learning for unseen-class unlabeled data

    Lan-Zhe Guo, Zhenyu Zhang, Yuan Jiang, Yufeng Li, and Zhi-Hua Zhou. Safe deep semi-supervised learning for unseen-class unlabeled data. In Proc. International Conference on Machine Learning (ICML) , pages 3897–3906, 2020

    2020

  11. [19]

    Binary decomposition: A problem transformation perspective for open-set semi-supervised learning

    Jun-Yi Hang and Min-Ling Zhang. Binary decomposition: A problem transformation perspective for open-set semi-supervised learning. In Proc. International Conference on Machine Learning (ICML) , pages 17505–17518, 2024

    2024

  12. [20]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 770–778, 2016

    2016

  13. [21]

    Safe-student for safe deep semi-supervised learning with unseen-class unlabeled data

    Rundong He, Zhongyi Han, Xiankai Lu, and Yilong Yin. Safe-student for safe deep semi-supervised learning with unseen-class unlabeled data. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 14565–14574, 2022

    2022

  14. [22]

    SAFER- STUDENT for safe deep semi-supervised learning with unseen-class unlabeled data

    Rundong He, Zhongyi Han, Xiankai Lu, and Yilong Yin. SAFER- STUDENT for safe deep semi-supervised learning with unseen-class unlabeled data. IEEE Trans. on Knowledge and Data Engineering, pages 318–334, 2024

    2024

  15. [23]

    Using self-supervised learning can improve model robustness and un- certainty

    Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and un- certainty. In Advances in Neural Information Processing Systems (NeurIPS), pages 15637–15648, 2019

    2019

  16. [24]

    Trash to treasure: Harvesting OOD data with cross-modal matching for open-set semi-supervised learning

    Junkai Huang, Chaowei Fang, Weikai Chen, Zhenhua Chai, Xiaolin Wei, Pengxu Wei, Liang Lin, and Guanbin Li. Trash to treasure: Harvesting OOD data with cross-modal matching for open-set semi-supervised learning. In Proc. IEEE International Conference on Computer Vision (ICCV), ...

    2021

  17. [25]

    They are not completely use- less: Towards recycling transferable unlabeled data for class-mismatched semi-supervised learning

    Zhuo Huang, Jian Yang, and Chen Gong. They are not completely use- less: Towards recycling transferable unlabeled data for class-mismatched semi-supervised learning. IEEE Trans. on Multimedia , pages 1844– 1857, 2022

    2022

  18. [26]

    Label propagation for deep semi-supervised learning

    Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Label propagation for deep semi-supervised learning. In Proc. IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR) , pages 5070–5079, 2019

    2019

  19. [27]

    Unknown-aware graph regularization for robust semi-supervised learn- ing from uncurated data

    Heejo Kong, Suneung Kim, Ho-Joong Kim, and Seong-Whan Lee. Unknown-aware graph regularization for robust semi-supervised learn- ing from uncurated data. In Proc. AAAI Conference on Artificial Intelligence (AAAI), pages 13265–13273, 2024

    2024

  20. [28]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Department of Computer Science, University of Tronto, 2009

    2009

  21. [29]

    Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper R. R. Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 14 Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images d...

    2021

  22. [30]

    Temporal ensembling for semi-supervised learning

    Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In Proc. International Conference on Learning Representations (ICLR), 2017

    2017

  23. [31]

    Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks

    Dong-Hyun Lee. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Proc. International Conference on Machine Learning (ICML) Workshops , page 896, 2013

    2013

  24. [32]

    Diversify and disam- biguate: Learning from underspecified data

    Yoonho Lee, Huaxiu Yao, and Chelsea Finn. Diversify and disam- biguate: Learning from underspecified data. In Proc. International Conference on Machine Learning (ICML) Workshops , 2022

    2022

  25. [33]

    Diversify and dis- ambiguate: Out-of-distribution robustness via disagreement

    Yoonho Lee, Huaxiu Yao, and Chelsea Finn. Diversify and dis- ambiguate: Out-of-distribution robustness via disagreement. In Proc. International Conference on Learning Representations (ICLR) , 2023

    2023

  26. [34]

    Junnan Li, Caiming Xiong, and Steven C. H. Hoi. Comatch: Semi- supervised learning with contrastive graph regularization. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 9455–9464, 2021

    2021

  27. [35]

    Iomatch: Simpli- fying open-set semi-supervised learning with joint inliers and outliers utilization

    Zekun Li, Lei Qi, Yinghuan Shi, and Yang Gao. Iomatch: Simpli- fying open-set semi-supervised learning with joint inliers and outliers utilization. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 15824–15833, 2023

    2023

  28. [36]

    Rethinking safe semi-supervised learning: Transferring the open- set problem to a close-set one

    Qiankun Ma, Jiyao Gao, Bo Zhan, Yunpeng Guo, Jiliu Zhou, and Yan Wang. Rethinking safe semi-supervised learning: Transferring the open- set problem to a close-set one. In Proc. IEEE International Conference on Computer Vision (ICCV) , pages 16370–16379, 2023

    2023

  29. [37]

    Umap: Uniform manifold approximation and projection for dimension reduction

    Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018

    2018 arXiv

  30. [38]

    Goodfellow

    Avital Oliver, Augustus Odena, Colin Raffel, Ekin Dogus Cubuk, and Ian J. Goodfellow. Realistic evaluation of deep semi-supervised learning algorithms. In Advances in Neural Information Processing Systems (NeurIPS), pages 3239–3250, 2018

    2018

  31. [39]

    A threshold selection method from gray-level his- tograms

    Nobuyuki Otsu. A threshold selection method from gray-level his- tograms. IEEE Trans. on Systems, Man, and Cybernetics , pages 62–66, 1979

    1979

  32. [40]

    Openmatch: Open-set consistency regularization for semi-supervised learning with outliers

    Kuniaki Saito, Donghyun Kim, and Kate Saenko. Openmatch: Open-set consistency regularization for semi-supervised learning with outliers. In Advances in Neural Information Processing Systems (NeurIPS) , 2021

    2021

  33. [41]

    Regulariza- tion with stochastic transformations and perturbations for deep semi- supervised learning

    Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regulariza- tion with stochastic transformations and perturbations for deep semi- supervised learning. In Advances in Neural Information Processing Systems (NeurIPS), pages 1163–1171, 2016

    2016

  34. [42]

    Fixmatch: Simplifying semi-supervised learning with consis- tency and confidence

    Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun- Liang Li. Fixmatch: Simplifying semi-supervised learning with consis- tency and confidence. In Advances in Neural Information Processing Systems (...

    2020

  35. [43]

    Graph-based semi-supervised learning: A comprehensive review

    Zixing Song, Xiangli Yang, Zenglin Xu, and Irwin King. Graph-based semi-supervised learning: A comprehensive review. IEEE Trans. on Neural Networks and Learning Systems , pages 8174–8194, 2023

    2023

  36. [44]

    The semi-supervised inaturalist challenge at the fgvc8 workshop

    Jong-Chyi Su and Subhransu Maji. The semi-supervised inaturalist challenge at the fgvc8 workshop. In arXiv preprint arXiv:2106.01364 , 2021

    2021 arXiv

  37. [45]

    Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results

    Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Advances in Neural Information Processing Systems (NeurIPS), pages 1195–1204, 2017

    2017

  38. [46]

    USB: A unified semi-supervised learning benchmark for classification

    Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, Heli Qi, Zhen Wu, Yufeng Li, Satoshi Nakamura, Wei Ye, Marios Savvides, Bhiksha Raj, Takahiro Shinozaki, Bernt Schiele, Jindong Wang, Xing Xie, and Yue Zhang. USB: A ...

    2022

  39. [47]

    Freematch: Self-adaptive thresholding for semi-supervised learning

    Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, Bernt Schiele, and Xing Xie. Freematch: Self-adaptive thresholding for semi-supervised learning. In Proc. International Conference on Learning Repre...

    2023

  40. [48]

    Out-of-distributed semantic pruning for robust semi-supervised learning

    Yu Wang, Pengchong Qiao, Chang Liu, Guoli Song, Xiawu Zheng, and Jie Chen. Out-of-distributed semantic pruning for robust semi-supervised learning. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 23849–23858, 2023

    2023

  41. [49]

    Scomatch: Alleviating overtrusting in open-set semi-supervised learning

    Zerun Wang, Liuyu Xiang, Lang Huang, Jiafeng Mao, Ling Xiao, and Toshihiko Yamasaki. Scomatch: Alleviating overtrusting in open-set semi-supervised learning. In Proc. European Conference on Computer Vision (ECCV), pages 217–233, 2025

    2025

  42. [50]

    Hovy, Thang Luong, and Quoc Le

    Qizhe Xie, Zihang Dai, Eduard H. Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. In Advances in Neural Information Processing Systems (NeurIPS) , 2020

    2020

  43. [51]

    Self-training for class- incremental semantic segmentation

    Lu Yu, Xialei Liu, and Joost van de Weijer. Self-training for class- incremental semantic segmentation. IEEE Trans. on Neural Networks and Learning Systems , pages 9116–9127, 2023

    2023

  44. [52]

    Multi-task curriculum framework for open-set semi-supervised learning

    Qing Yu, Daiki Ikami, Go Irie, and Kiyoharu Aizawa. Multi-task curriculum framework for open-set semi-supervised learning. In Proc. European Conference on Computer Vision (ECCV) , pages 438–454, 2020

    2020

  45. [53]

    Wide residual networks

    Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Proc. British Machine Vision Conference (BMVC) , 2016

    2016

  46. [54]

    Flexmatch: Boosting semi- supervised learning with curriculum pseudo labeling

    Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi- supervised learning with curriculum pseudo labeling. In Advances in Neural Information Processing Systems (NeurIPS) , 2021

    2021

  47. [55]

    Simmatchv2: Semi-supervised learning with graph consistency

    Mingkai Zheng, Shan You, Lang Huang, Chen Luo, Fei Wang, Chen Qian, and Chang Xu. Simmatchv2: Semi-supervised learning with graph consistency. In Proc. IEEE International Conference on Computer Vision (ICCV), pages 16386–16396, 2023

    2023

  48. [56]

    Simmatch: Semi-supervised learning with similarity match- ing

    Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity match- ing. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 14451–14461, 2022. Heejo Kong received the B.S. degree in Ma...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.