acceptodds
Withdrawn from ICLR 2027

Strong Views for Representation, Weak Views for Classification: Rethinking Augmentation in Class-Imbalanced Semi-Supervised Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Class-imbalanced semi-supervised learning (CISSL) aims to leverage unlabeled data for both effective representation learning and balanced classification. Existing methods commonly use weak-to-strong augmentation consistency for both objectives, but we show that the appropriate augmentation strength should instead be chosen according to the learning objective. Strong augmentation promotes invariant representation learning, but it also shifts class-conditional means toward their center in feature space substantially more than weak augmentation, reducing the distance between classes. Under class imbalance, these smaller distances amplify the imbalance-induced boundary shift toward tail classes, shrinking their decision regions. These findings motivate a role-separated principle: use strong augmentation to learn representations, but weak augmentation to estimate the decision boundary. We instantiate this principle with the Weak-View Classifier (WVC), which estimates the final decision boundary from weakly augmented samples, where the distances between class means are better preserved. To improve the pseudo-labels used for decision-boundary estimation, we further introduce Reference-Guided Logit Adjustment (RGLA), which uses class prototypes as a reference to adapt the strength of logit adjustment without knowledge of the unlabeled class distribution. Across four CISSL benchmarks and multiple base SSL algorithms, our approach consistently improves decision-boundary estimation and achieves competitive or superior performance across diverse settings.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Class-imbalanced semi-supervised learning, Strong - and weak augmentation, Logit-adjustment

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 3 reviews, above 27% of the others.
Reviewer_1qLmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes the application of separation enhancement in the context of class-imbalanced semi-supervised learning: The encoder continues to learn representations with consistency from weak to strong, while the independent linear Weak-View Classifier (WVC) only uses weakly enhanced features to train the final classification boundary, and its loss is not backpropagated to the encoder. Reference-Guided Logit Adjustment (RGLA) uses the cosine nearest class mean classifier as a reference, and dynamically updates the logit adjustment strength of the backbone network based on the difference in head category preferences predicted by the two branches. The authors illustrate the effect of reducing the class mean distance on amplifying the boundary displacement caused by prior imbalance using an equal-variance two-class Gaussian model, and provide a comparison of the weak/strong views of the encoder, contraction of the class means, and data balancing control to support this explanation. Competitively good results are reported on four benchmark classes and multiple SSL backbones; the appendix also provides the weak view transfer of ABC, the compatibility of FreeMatch/SoftMatch, and the analysis of the controller. The theory here is to simplify the mechanism analysis, and not to provide a complete guarantee for the entire pseudo-label and adjustment loss training process.

Critical strengths
  1. The core enhancement principle has strong control experiments rather than relying solely on final accuracy. Table 6 only changes the view used for the unlabeled classification loss on the same encoder, and the weak view increases by 1.1–4.4 percentage points in six settings. Table 1's synthetic contraction maintains the relative positions within the pseudo-label group, and the tail class recall decreases as the contraction increases; under the balanced condition, the weak/strong difference basically disappears, and the additional experiments of ABC retain the original weight balancing mechanism. In summary, these evidences effectively support the separation of the enhancement role, rather than mixing all benefits in a new system.
  2. The method is lightweight and has practical usability on multiple distribution conditions and backbone networks. WVC does not modify the representation learning objective of the encoder, and RGLA does not require the actual proportion of unlabeled classes; results for matching, mismatching, and reverse distribution settings, as well as on four types of datasets, cover the most important usage scenarios of the method. The gap with the strong baseline is small, which cannot be simply explained as a significant victory, but combined with the fixed τ, prototype sample size, control parameters, and running cost checks in the appendix, the competitiveness and ease of integration have credible support.
Critical weaknesses
  1. The Boundary Amplification Theorem does not directly correspond to the actual used prior-adjusted classifier. Proposition 1 analyzes the unadjusted Bayes boundary with training prior, while WVC adds log πall to the training loss and removes this bias during testing. According to the loss in this paper, the unbiased test logit boundary is b = (σ²/d)[log(p_h/p_t) - log(a_h/a_t)]. When a = p, this prior shift disappears. Therefore, the weak view advantage of the actual WVC still requires residual prior error, pseudo-label errors, finite samples, or variance changes, etc. to connect; the existing proposition itself cannot complete this step. This problem limitation mechanism explains the completeness of the interpretation, but does not negate the existing weak/strong contrast results. (Pages 4, 6–7, Equations 5–7, Proposition 1.)
  2. The sample number weights of the combined prior and the written optimization objective may not be consistent. Equation 5 constructs πall = (Nπ + Mqback) / (N + M) based on the dataset size N and M, but Equations 6 and 7 respectively calculate the mean based on the labeled and unlabeled minibatch sizes, and then add them equally. If each source is uniformly sampled and indeed optimized according to this objective, the corresponding category mixing weights should be determined by the relative coefficients of the two losses; in the case of equal weights, it is (π + qback) / 2, rather than N:M. When the two distributions match, the difference is not apparent, but when they do not match, it may change the adjustment direction and size. It is necessary to explain the loss/sampling weights in the implementation, or to position Equation 5 as an empirical prior selection rather than a natural statistical correction derived from Equations 6–7. (Page 4, Equations 5–7; Page 5, Equation 10.)
Questions for authors
  1. Are there any unrecorded N:M loss weights, non-uniform source sampling, or other compensations that make Equation 5 consistent with Equations 6–7? If not, how would constructing a prior based on the effective loss weights affect the weak/strong view gap and the final accuracy in a mismatched setting? This will determine whether the second question is merely a statement omission or a real method choice.
  2. Please explain how Proposition 1 covers the test boundaries after using logit adjustment. In the existing experiments, which residual imbalance or pseudo-label bias makes the displacement term still exist? Provide corresponding limiting conditions, or separate the "adjusting the prior accuracy/inaccuracy" scenario in the existing fixed encoder setting, which can significantly strengthen the mechanism conclusion.

more in Weaknesses

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

The paper studies augmentation choices in class-imbalanced semi-supervised learning. It trains a separate Weak-View Classifier on detached weak-view features while retaining strong augmentation for backbone representation learning. Reference-Guided Logit Adjustment uses a prototype classifier to adapt the backbone's adjustment strength. The authors explain the weak-view advantage through class-mean contraction and imbalance-induced boundary shifts. They evaluate the method on four image-classification benchmarks, with a Gaussian analysis, feature interventions, and ablation studies.

Critical strengths

Controlled experiments. Using the same trained encoder, Table 6 shows that weak-view classifier training outperforms its strong-view counterpart in all six CIFAR-10-LT settings. Appendix H finds a similar advantage with ABC's undersampling scheme, while Appendix G finds little difference under balanced class frequencies. These comparisons provide consistent evidence for the proposed augmentation choice under class imbalance and help separate its benefit from differences in the learned representation.

Critical weaknesses
  1. Modest gains over LLA. Across the twelve matched-imbalance bACC comparisons in Table 2, WVC improves over LLA by only 0–0.8 percentage points. On CIFAR-10-LT with FixMatch, it matches LLA at γ=50\gamma=50 and improves by 0.2 points at γ=100\gamma=100 and 150150. I find the empirical advance over LLA limited in these matched-distribution settings.

  2. Difficult presentation. The roles of the backbone classifier, WVC, and cosNCM are difficult to track through the notation for pseudo-labels, EMA marginals, and logit offsets. Following how the classifier branches and controller interact requires repeatedly moving between the equations and Algorithm 1.

Questions for authors
  1. Correction prior: Eq. (5) weights the two datasets by N:MN:M, while Eqs. (6)–(7) average each source's loss separately and add them with equal weight. With a fixed pseudo-label marginal qq and uniform sampling within each source, the two equally weighted source distributions have label marginal (π+q)/2(\pi+q)/2. How does the count-weighted adjustment prior (Nπ+Mq)/(N+M)(N\pi+Mq)/(N+M) account for these equal source weights in the optimized objective?

  2. Adjusted boundary: Proposition 1 analyzes the ordinary Bayes boundary, whereas WVC uses unit-strength logit adjustment during training and raw logits at test time. Appendix H attributes the weak-view advantage under undersampling to amplification of a residual offset. What residual offset remains under WVC's logit-adjusted objective, and is its amplification measured in the class-mean intervention in Table 1?

  3. Class-mean distances: Eq. (11) measures movement toward the centroid along a particular direction. Do direct measurements of pairwise class-mean distances also show that strong augmentation brings class means closer together in the tested settings?

  4. Intervention control: The intervention in Sec. 5.3 preserves covariance within each pseudo-label group. Were true-class conditional covariances also measured before and after the intervention, and do they remain approximately unchanged despite pseudo-label errors?

  5. Baseline comparability: Table 4 reports reversed-setting bACC of 74.074.0 for FixMatch+LLA and 77.577.5 for ReMixMatch+LLA, versus 78.378.3 and 82.482.4 in LLA's published Table 2. Are these reruns, and which settings explain the differences? For Table 5, were baselines rerun with WRN-28-2 or taken from their original papers? Sec. C.2 states WRN-28-2 for all experiments, while LLA's original setup uses ResNet-50 for Small-ImageNet-127.

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

The paper argues that in class-imbalanced SSL, strong augmentation should be used for representation learning but not for decision-boundary estimation. Strong augmentation contracts class means toward the centroid (wˉ\bar{w} ≈ 0.52-0.60 vs. 0.04-0.09 for weak). A two-Gaussian analysis (Prop. 1) suggests this amplifies the boundary shift toward tail classes. The authors propose WVC, a detached linear head trained on weak views, and RGLA, a controller that adapts logit-adjustment strength τ\tau by matching the backbone's head-preference to a cosNCM reference.

Critical strengths
  1. Clear, testable hypothesis with a measurement, theory, intervention, control chain.
  2. Informative controls: the balanced-data result (Tab. 9-10) and the ABC view-swap (Tab. 11).
  3. Cheap method (negligible overhead) evaluated across matched, mismatched, reversed and unknown-γu settings.
Critical weaknesses
  1. Theory vs. experiments. Prop. 1 assumes no prior correction, but WVC applies logit adjustment (λ^all). If the adjustment were accurate, the gap should vanish. The "residual offset" is never measured.
  2. Confounds. Within-class variance also rises under strong augmentation, so d/σ matters, not d. No per-view Fisher/d′, no head-vs-tail ω_k, and train/test view shift is not ruled out. Tab. 1 is a single setting without variance.
  3. Tab. 6 baseline. At (100,1), the strong-view classifier (85.2) is far below the RGLA backbone (89.3) and WVC (89.6), so the gain over the backbone is about +0.3, not +4.4. The protocol is unspecified, and backbone rows and a full ablation (backbone, + WVC, + RGLA) are missing.
  4. Small SOTA gains. On CIFAR-10-LT, gains over LLA are within one std with three seeds, and RGLA's sensitivity (~0.6) is comparable. Tab. 4, 5, 19 and 20 lack std. Baseline provenance and hyperparameter selection (no validation split) are unclear.
  5. RGLA justification and scope. The cosNCM reference is biased and co-evolves with the encoder. No convergence argument is given, and Eq. 5 estimates the unlabeled distribution despite the claim of not needing it. The "strong views for representation" half is untested, with no augmentation-strength sweep.
Questions for authors
  1. Does the strong-view classifier use the same λ^all, pseudo-labels and RGLA? What are the backbone numbers in Tab. 6?
  2. How do the gap and tail recall vary with augmentation strength?
  3. What residual boundary offset remains after logit adjustment?
  4. Are the improvements over LLA statistically significant, and were baselines re-run?

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

Related papers

Discussion (0)

Sign in to comment.