- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes Multi-Source Multi-View (MSMV), a contrastive self-supervised learning framework for long-tailed data. MSMV constructs positive pairs by fusing complementary view configurations of two randomly paired images using online and momentum encoders. It additionally introduces a filtered contrastive loss (FCL) that modifies similarities outside a specified interval. The motivation is that cross-image interactions can alleviate frequency-dependent optimization bias and improve tail-class representations. Evaluation includes linear classification on CIFAR-10/100-LT and ImageNet long-tailed settings, gradient-ratio diagnostics, augmentation and loss ablations, and representation analyses. The reported results show improvements over the evaluated baselines, particularly on CIFAR-100-LT.
Critical strengths
-
Promising empirical improvements on a relevant problem. Learning useful representations from unlabeled long-tailed data is practically important. On CIFAR-100-LT, MSMV+ achieves 60.18% overall accuracy and 61.50% rare-class accuracy, compared with 55.27% and 53.55% for FASSL. These reported improvements suggest that the proposed combination merits further investigation.
-
The evaluation goes beyond aggregate accuracy. The paper includes head/medium/tail results, gradient diagnostics, augmentation comparisons, and ablations of FCL. In particular, the loss ablations provide useful evidence that FCL contributes to the performance of the combined framework.
Critical weaknesses
-
The central mechanism connecting feature fusion to reduced long-tail bias is not established. Section 3 motivates the method through a shared-component estimation example, while Appendix .11 argues that fusion introduces an additional gradient pathway. However, neither establishes that this pathway transfers useful complementary semantics, expands tail-class semantic coverage, or reduces frequency-dependent bias. Standard contrastive learning already shares encoder parameters across samples, and its gradients depend on other examples through the loss. Consequently, the existence of an additional dependency does not establish the claimed distinction in learning behavior. Likewise, the shared-component example is not connected mathematically to the actual fusion objective. The gradient-ratio results are suggestive, but do not establish that a ratio near 0.5 is desirable or that deviations are the root cause of representation degradation. The current evidence remains compatible with alternative explanations such as feature augmentation, regularization, or changes in positive-pair difficulty. This substantially weakens the paper’s main conceptual contribution, although it does not negate the reported empirical gains.
-
The experimental comparisons do not sufficiently establish the necessity of the proposed design or the breadth of its claimed advantage. Figure 5 and the FCL ablations are useful, but do not isolate the proposed mechanism from simpler alternatives. For example, the paper does not report a controlled comparison between its crossed fusion,
(A_n + C_a) ↔ (A_a + C_n), and(A_n + C_n) ↔ (A_a + C_a), using the same four views. Nor does it compare against feature-noise controls or establish how much of the benefit can be obtained by adding FCL to the underlying baseline. These comparisons matter because they could change the interpretation of the contribution. Baseline coverage is also limited: the newest method in the main tables is FASSL (2023), ImageNet comparisons omit relevant long-tailed SSL methods such as TS and Geometric Harmonization, and the subsampling experiments compare only against MoCoV2. Thus, the large gains in the latter setting do not establish superiority over strong specialized alternatives. Additionally, proportional subsampling primarily tests reduced data availability rather than increased class imbalance, and the claimed 8% “few-shot” improvement lacks an identifiable few-shot evaluation protocol.
Questions for authors
-
Mechanism clarification: What evidence specifically supports the claim that fusion transfers useful complementary semantics to tail classes, rather than acting as feature augmentation or regularization? Is there a result linking the additional gradient pathway to reduced frequency sensitivity, beyond the observed gradient ratios and downstream accuracy?
-
Discriminating evidence: Do you have controlled results for
(A_n + C_n) ↔ (A_a + C_a), baseline + FCL, or feature perturbations matched in scale and, where possible, cross-view correlation to the counterpart features? These are not requests for a broad future-work agenda: such evidence would materially affect whether I view the contribution as a long-tail-specific mechanism or an effective augmentation/filtering combination. -
Loss implementation: Equation (7) appears to set out-of-range similarities to zero before exponentiation, so those entries still contribute
exp(0) = 1to the denominator. Are these entries actually zeroed or excluded in the implementation, and does filtering apply to positive pairs as well? Please clarify the exact objective used for the reported experiments. -
Evaluation clarification: Does the reported 8% “few-shot” gain refer to the rare-class accuracy difference in Table 2, or to a separate few-shot protocol? Also, are comparisons with TS or Geometric Harmonization available under the same ImageNet training and evaluation settings? These clarifications would affect my assessment of the scope and strength of the empirical claims.