Pith. sign in

REVIEW 3 major objections 3 minor 22 references

Weakly-Supervised Contrastive Learning for Imprecise Class Labels

T0 review · 3 major / 3 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 The paper proves that a weakly-supervised contrastive loss, built on estimated semantic similarity between examples, can approximate supervised contrastive learning, with linear-probe error controlled by the bias of the similarity estimate.

desk verdict A genuinely new spectral-graph derivation of a weakly-supervised contrastive loss with strong empirical gains, but the headline bound is conditional on an unmeasured bias term that the self-labeling scheme does not control. read the letter →

arxiv 2505.22028 v1 pith:TWHLH5EF submitted 2025-05-28 cs.LG

Zi-Hao Zhou , Jun-Jie Wang , Tong Wei , Min-Ling Zhang This is my paper
classification cs.LG
keywords weakly-supervisedcontrastivelearningcontinuoussemanticsimilaritynoisylabelpartialgraphspectralclusteringperturbationaugmentationlinearprobeerrorbound
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

This paper claims that contrastive representation learning can work with imprecise labels—noisy labels or partial labels—by replacing the binary decision "same class or not" with a continuous semantic similarity between examples. It builds a graph whose vertices are augmented data points and whose edge weights mix self-supervised connectivity with this semantic similarity, then shows that minimizing the resulting weakly-supervised contrastive loss is equivalent to spectral clustering of that graph. The paper proves that, when an estimated similarity matrix $S$ makes $S(x)P(q|x)$ approximate the true class posterior $P(y|x)$, the linear-probe error of the learned representation is bounded by the clustering error of the graph plus finite-sample error plus the bias $\Delta(S)$. Experiments with noisy labels and partial labels report consistent gains, with the largest improvements at the highest noise and partial-label rates. The theoretical point is that weakly-supervised contrastive learning can approximate fully supervised contrastive learning as the semantic similarity estimate becomes accurate.

What carries the argument

The load-bearing identity is Proposition 2.1: for any recovery matrix $S$ with $P(y|x)=S(x)P(q|x)$, the weakly-supervised edge weight $w^{wl}(S)$ coincides with the supervised edge weight $w^l$. The paper constructs a perturbation augmentation graph with adjacency $A = \alpha A^u + \beta A^{wl}(S)$, where $A^u$ is the self-supervised augmentation graph; Proposition 2.3 shows the weakly-supervised contrastive loss $L_{wsc}$ equals the Frobenius-norm spectral clustering objective on that graph, so minimizing the loss recovers the top eigenvectors of the normalized adjacency matrix. The named object is the semantic similarity matrix $S$ (and its estimate $\hat{S}$), which converts weak supervision $q$ into continuous edge weights; its bias $\Delta(\hat{S})$ is the term that separates weakly-supervised from supervised performance in the final bound.

What would settle it

On a dataset with known true labels, such as CIFAR-10 or CIFAR-100 with 90% symmetric noise, track the empirical bias $\Delta(\hat{S}) = E[\|P(y|x) - \hat{S}(x)P(q|x)\|_1]$ during training and compare it with the linear-probe error curve; if the probe error drops substantially while $\Delta(\hat{S})$ stays large, the bias term in the bound is not the mechanism driving the gain. A stronger control is to replace $\hat{S}$ with a fixed random matrix matching only the class prior and check whether the weakly-supervised loss still improves over the self-supervised baseline.

Watch

Extended reading notes

Core claim

The central discovery is that supervised contrastive edge weights $w^l_{x,x'}$ can be reproduced from weak supervision alone. If $S: X \to \mathbb{R}^{c \times v}$ satisfies $P(y|x)=S(x)P(q|x)$ almost everywhere, then the weakly-supervised edge weight $w^{wl}(S)$ equals $w^l$, so the semantic similarity $S((\tilde{x},\tilde{q}),(\tilde{x}',\tilde{q}')) = S(\tilde{x})^T_{:,\tilde{q}} S(\tilde{x}')_{:,\tilde{q}'}$ is a continuous stand-in for the indicator $I[\tilde{y}=\tilde{y}']$. The paper defines a perturbation augmentation graph $A = \alpha A^u + \beta A^{wl}(S)$, proves that the weakly-supervised contrastive loss $L_{wsc}$ equals the spectral clustering objective $\|\tilde{A} - F F^T\|_F^2$ up to a constant, and derives Corollary 3.8: for representation dimension $d > 4r$ and $\alpha + \beta/c = 1$, with high probability the linear-probe error is $\tilde{O}$ of the graph clustering term plus finite-sample Rademacher terms plus $\beta \Delta(\hat{S})$, where $\Delta(\hat{S}) = E[\|P(y|x) - \hat{S}(x)P(q|x)\|_1]$. The bound makes precise a trade-off: increasing $\beta$ improves the graph's intra-class density but also amplifies the bias and variance of the estimated similarity matrix.

Load-bearing premise

The whole guarantee rests on having an estimated matrix $S(x)$ such that $S(x)P(q|x)$ is close to the true class distribution $P(y|x)$; in the main experiments $S(x)$ is just the model's own current predictions, and the theory does not guarantee those predictions are accurate.

Editorial extensions

If this is right

  • If the bound is right, any weakly-supervised learner that can produce a low-bias matrix $\hat{S}$ gets a contrastive regularizer that provably approaches supervised contrastive learning as $\beta$ and sample size grow.
  • The framework applies to noisy labels and partial labels through the same $S$-construction recipe, and since only $S((\tilde{x},\tilde{q}),(\tilde{x}',\tilde{q}'))$ enters the loss, the same mechanism extends to other weak-supervision formats.
  • The bias-variance split predicted by Theorem 3.7 implies that combining environment information (e.g., transition-matrix estimates) with sample information (model predictions) should beat either alone; the paper's ablations confirm this ordering.
  • Performance gains should be largest when weak supervision is most ambiguous, matching the reported improvements of roughly 6.8% at 90% noise and 7.8% at a partial ratio of 0.3.

Reading between the lines

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

  • A direct test the paper does not run: use an oracle $S$ built from true labels and compare linear-probe error against the self-labeled $\hat{S}$; the difference would isolate how much of the empirical gain is the framework itself versus the prediction bootstrap.
  • The variance term $\sup_x \|\hat{S}(x)^T \hat{S}(x)\|_\infty$ suggests that ensembling or sharpening prediction-based $\hat{S}$ should improve the bound without changing the loss, a cheap extension for practitioners.
  • The same graph-theoretic construction predicts that the method will help whenever weak supervision is informative about class co-membership, such as bag-level or multi-modal supervision, and will fail when the weak signal is pure noise, because the framework converts weak-supervision quality directly into edge-weight quality.
  • One could measure $\Delta(\hat{S})$ on real noisy-label datasets such as CIFAR-10N and check whether the bias term tracks the reported accuracy ordering across noise types (aggregate vs worst); the paper does not report this diagnostic.
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

3 major / 3 minor

Summary. This paper introduces weakly-supervised contrastive learning (WSC), a graph-spectral framework in which the discrete positive/negative pair structure of supervised contrastive learning is replaced by continuous semantic similarity weights estimated from weak labels (noisy or partial). The authors derive an exact equivalence between a spectral matrix-factorization objective and a tractable contrastive loss (Prop. 2.3), propose constructions of the similarity matrix for noisy-label and partial-label settings, prove a downstream linear-probe bound (Cor. 3.8) with terms for graph clustering, finite samples, and a bias term, and report accuracy gains over existing NLL/PLL baselines across CIFAR, CUB-200, Clothing1M, and instance-dependent PLL datasets.

Significance. The elegant derivation of Prop. 2.3 and the modular treatment of weak supervision as a perturbation of the augmentation graph are genuine contributions. If the main bound were matched to the evaluated algorithm, the paper would provide a principled explanation for the reported gains. The experimental scope is broad and includes ablations, CIFAR-N/Clothing1M, and IDPLL, and the release of code is a strength. The main limitation is that the central guarantee is conditional on a bias term that is neither bounded nor measured for the self-labeling construction used in the headline experiments; the current evidence therefore does not conclusively support the claim that WSC approximates supervised contrastive learning in practice.

major comments (3)
  1. [§3.2 / Def. 3.6 / Cor. 3.8 / App. F.1] The main bound (Eq. 29) contains the additive term βη'_2Δ(Ŝ), where Δ(Ŝ)=E_x||P(y|x)−Ŝ(x)P(q|x)||_1. In the main experiments, Ŝ is constructed from the current network's predictions: Ŝ(x)_{y,q}=g(A_w(x))_y for NLL and Ŝ(x)_{y,q}=I[y∈q]g(A_w(x))_y / Σ_{y'∈q}g(A_w(x))_{y'} for PLL (Appendix F.1, "Construct of Main Experiments"). Substituting these constructions into Def. 3.6 makes Δ(Ŝ) essentially the model's own error or miscalibration on weakly augmented inputs. No theorem in the paper bounds this quantity, and Theorem 3.7 only carries it linearly into the final error bound; moreover, because Ŝ is re-estimated from the same network being trained, the empirical gains in Tables 1 and 2 could in principle be produced by the cross-entropy or consistency terms of Eq. (112) once g(A_w(x)) becomes informative. Table 10 shows that adding the WSC term helps, but it does not measure Δ(Ŝ), so it cannot separate a genuine semantic-similarity signal from a self-training bootstrap. I ask the authors to either prove smallness of Δ(Ŝ) for a concrete Ŝ, measure it during the training runs of Tables 1 and 2, or explicitly restrict the theoretical claims to constructions for which the bias is controlled.
  2. [Def. 3.1 / Lemma C.3] Definition 3.1 defines γ-consistency as E_{(˜x,˜y),x∼A}[I[˜y=ˆy(x)]] ≤ γ. Taken literally, this says the pseudo-labeler agrees with the true label with probability at most γ, so a smaller γ corresponds to a worse labeler, contradicting the interpretation in Theorem 3.4 that small γ implies sparsity of connections between classes. The proof of Lemma C.3 (Eq. 64) uses the complementary event I[ˆy(x)≠˜y] in the expected term. The definition should be corrected to E[I[˜y≠ˆy(x)]] ≤ γ, and all statements involving γ should be checked with this corrected definition.
  3. [E.1 / Eq. 14 / Eq. 112] The CIFAR-100 experiments use α=2, β=300 and additionally scale the last term of the loss by a factor of 3. With c=100, α+β/c=5, so the implemented loss is not the loss of Eq. (14), whose L3 coefficient would be (α+β/c)^2 = 25; the extra factor of 3 cannot be absorbed into α and β because the coefficients of L1 and L2 are fixed by Eq. (14). Theorem 3.7 and Corollary 3.8 assume α+β/c=1, so the theoretical results do not directly cover the CIFAR-100 experiments. Please clarify whether the scaling is an implementation detail outside the theory or adapt the analysis to the actual loss used.
minor comments (3)
  1. [Cor. 3.8 / App. C.3] The statement of Corollary 3.8 uses d > 4r but r is not defined in the main text; the restated Theorem D.7 requires d > 4r+2 and defines Δλ. Please define r and align the dimensional assumptions.
  2. [Algorithm 1 / Def. C.10] Algorithm 1 uses B_Q and B_U for batch sizes, while the empirical loss in Definition C.10 is written with n_q and n_u; the connection between the batch estimator and the population loss should be spelled out, especially the normalization of the estimated L2 term by B_Q^2.
  3. [Proceedings line] The proceedings line reads "42 st International Conference on Machine Learning" and should read "42nd."

Circularity Check

1 steps flagged · score 4.0 of 10

The formal spectral derivation is non-circular, but the evaluated algorithm is partially circular: the semantic-similarity matrix is generated by the very model being trained, and Corollary 3.8's bias term reduces to that model's own unmeasured calibration error.

  1. fitted input called prediction [Appendix F.1 ('Construct of Main Experiments'), Algorithm 1, and Eq. (112)]
    "For the sake of simplicity in this experiment, the bS selected in the main experiments of this paper only adopts the simplest form... We simply take bS as bS(x)y,q = bP(y|x, q) =g(Aw(x))y for any x. ... Algorithm 1: Compute bL2 = ∥(S(X)T S(X))⊗(X1Q(X2Q)T )∥1 / B2Q."

    The semantic-similarity weights that define positive/negative pairs in L2 are S(X)^T S(X), and in the main experiments S(x)_{y,q}=g(A_w(x))_y, the current network's own posterior. The weak supervision entering the contrastive loss is therefore a function of the very model being trained; the graph is self-generated. Under this construction, the bias term in Corollary 3.8 is exactly Δ(bS)=E_x||P(y|x)-g(A_w(x))P(q|x)||_1, i.e., the model's own conditional-error/calibration quantity. No theorem bounds this quantity or proves that minimizing L_wsc drives it down; Theorem 3.7 only carries it linearly into the final bound. The empirical gains can thus be attributed to a self-training bootstrap plus the supervised and consistency terms in Eq.

full rationale

The derivation from the perturbation graph to the weakly-supervised contrastive loss (Proposition 2.3), the clustering error bound (Theorem 3.4), and the finite-sample Rademacher decomposition (Theorem 3.7) is self-contained algebra and does not reduce to its inputs. Corollary 3.8 is a conditional bound: it holds given a small bias Δ(bS) and small Rademacher terms, which is an honest statement rather than a hidden equivalence. The only load-bearing self-referential step is the empirical instantiation: Appendix F.1 sets bS to the current model's posterior g(A_w(x)), so the 'semantic similarity' graph is generated by the network under training. For that construction Δ(bS) is the model's own unmeasured calibration error, and the paper does not show that minimizing L_wsc reduces it. Thus the headline claim that the framework can approximate supervised contrastive learning is not established for the exact algorithm evaluated; the practical signal is partially a self-training bootstrap. Self-citations in the paper (e.g., Zhou et al. 2024, Wu et al. 2022/2024, Li et al. 2023) are background and are not load-bearing, and no uniqueness theorem from the authors is invoked. The central spectral argument retains independent content, so the circularity is partial rather than total.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The framework rests on hyperparameters α and β tuned per dataset, on the standard spectral-graph and Rademacher machinery, and on the unverified assumption that the self-labeling construction approximates the true posterior. The continuous semantic similarity is a new conceptual entity but not an independent physical one.

free parameters (3)
  • α (self-supervised coefficient) = 1 (CIFAR-10, CIFAR-10N, Clothing1M), 2 (CIFAR-100, CUB-200, fine-grained IDPLL)
    Controls the weight of the self-supervised spectral contrastive term L1 relative to the weakly-supervised term; tuned per dataset.
  • β (weakly-supervised coefficient) = 12 (CIFAR-10), 300 (CIFAR-100), 28 (Clothing1M), 6 (CIFAR-10N), 0-400 linear ramp (CUB-200, IDPLL)
    Controls the weight of the semantic-similarity term L2; the paper notes selecting an appropriate β is crucial (Remark after Cor 3.8).
  • Scaling factor on the final loss term = 3
    Applied to the last term of the loss for CIFAR-100 (Appendix E.1); an extra tuned hyperparameter not mentioned in the main text.
assumptions (5)
  • standard math Spectral graph theory results (Chung 1997, Louis-Makarychev 2014, Eckart-Young 1936), including higher-order Cheeger's inequality and the eigenvector-based characterization of optimal embeddings.
    Used in Lemma C.7, Lemma C.4, and Theorem 3.4 to translate graph conductance into linear probe error bounds.
  • domain assumption Augmentation distribution A(·|x) is such that the augmentation graph is γ-consistent (Definition 3.1): there exists a pseudo-labeler whose error rate against true class is at most γ.
    Invoked in Theorem 3.4 and its proof to upper-bound cross-class connections in the self-supervised graph.
  • domain assumption Uniform class distribution P(y)=[1/c,...,1/c]^T for Algorithm 1, Equation 14, and the proofs in Appendix C.
    The simplified loss and several equalities (e.g., L4=L3/c^2, L5=L3/c, degree normalization in Eq. 75) rely on uniform class prior. Experiments on CIFAR are balanced, but Clothing1M and CUB-200 are not strictly so.
  • domain assumption SCUI assumption (Assumption F.1) for partial-label transition estimation: candidate sets are generated independently and uniformly per false label.
    Used in Appendix F.2 to reduce PLL to c binary noisy-label problems and construct Ŝ via mixture proportion estimation.
  • ad hoc to paper The estimated matrix Ŝ approximately satisfies P(y|x) = Ŝ(x)P(q|x) (Prop 2.1 condition).
    The central identity w^{wl}(S)=w^l holds exactly only under this condition; the theory bounds the error by Δ(Ŝ), but no method is provided to ensure Δ(Ŝ) is small for the self-labeling construction used in the main experiments.
invented entities (1)
  • Continuous semantic similarity
    purpose: A scalar weight between 0 and 1 assigned to each pair of augmented samples, defined as S((x,q),(x',q')) = S(x)^T_{:,q} S(x')_{:,q'}; used as the edge weight in the perturbation augmentation graph to replace discrete positive/negative pairs.
    It is a conceptual construct introduced in this paper; its utility is demonstrated only through the paper's experiments and no external falsifiable handle is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Weakly-Supervised Contrastive Learning for Imprecise Class Labels." pith.science (2026). https://pith.science/paper/TWHLH5EF

@misc{pith2026250522028,
  author       = {Pith},
  title        = {Pith review of: Weakly-Supervised Contrastive Learning for Imprecise Class Labels},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TWHLH5EF}},
  note         = {Machine review of arXiv:2505.22028}
}
read the original abstract

Contrastive learning has achieved remarkable success in learning effective representations, with supervised contrastive learning often outperforming self-supervised approaches. However, in real-world scenarios, data annotations are often ambiguous or inaccurate, meaning that class labels may not reliably indicate whether two examples belong to the same class. This limitation restricts the applicability of supervised contrastive learning. To address this challenge, we introduce the concept of ``continuous semantic similarity'' to define positive and negative pairs. Instead of directly relying on imprecise class labels, we measure the semantic similarity between example pairs, which quantifies how closely they belong to the same category by iteratively refining weak supervisory signals. Based on this concept, we propose a graph-theoretic framework for weakly-supervised contrastive learning, where semantic similarity serves as the graph weights. Our framework is highly versatile and can be applied to many weakly-supervised learning scenarios. We demonstrate its effectiveness through experiments in two common settings, i.e., noisy label and partial label learning, where existing methods can be easily integrated to significantly improve performance. Theoretically, we establish an error bound for our approach, showing that it can approximate supervised contrastive learning under mild conditions. The implementation code is available at https://github.com/Speechless-10308/WSC.

Figures

Figures reproduced from arXiv: 2505.22028 by the authors.

Figure 1
Figure 1. (a) Self-supervised contrastive learning constructs positive example pairs by using different views of the same image and constructs negative example pairs by using different images. (b) Supervised contrastive learning further regards different images of the same class as additional positive example pairs. (c) Our proposed weakly-supervised contrastive learning abandons the concepts of discrete positive and negative… view at source ↗
Figure 2
Figure 2. We present t-SNE visualizations of the learned representations on the CIFAR dataset with 90% symmetric noise [PITH_FULL_IMAGE:figures/full_fig_p032_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 20 canonical work pages

  1. [1]

    and P(x|l i = 0). Under many assumptions to ensure identifiability, including the irreducibility assumption (Scott et al., 2013; Garg et al., 2021), the anchor point assumption (Scott, 2015; Liu and Tao, 2016), the separability assumption (Ramaswamy et al., 2016), θi can been estimated through many off-the-shelf methods (Scott et al., 2013; Scott, 2015; R...

    work page 2013

  2. [4]

    Table 4.Hyper-parameters fornoisy label learningused in experiments. Hyper-parameter CIFAR-10 (CIFAR-10N) CIFAR-100 (CIFAR-100N) Clothing1M Image Size 32 32 224 Model PreAct-ResNet-18 (ResNet-34) PreAct-ResNet-18 (ResNet-34) ResNet-50 (ImageNet-1K Pretrained) Batch Size 128 128 64 Learning Rate 0.02 0.02 0.002 Weight Decay 1e-3 1e-3 1e-3 LR Scheduler Cosi...

    work page 2020

  3. [5]

    Each runs has been repeated 3 times with different randomly-generated noise and we report the best accuracy

    Comparisons with each methods on CIFAR-10N, CIFAR-100N and Clothing1M. Each runs has been repeated 3 times with different randomly-generated noise and we report the best accuracy. Dataset CIFAR-10N CIFAR-100N Clothing1M Noisy Type Random1 Aggregate Worst Clean Noisy Ins. CE 85.02±0.6587.77±0.3877.69±1.55 76.70±0.7455.50±0.66 69.10 Forward 86.88±0.5088.24±...

    work page 2024

  4. [6]

    Lemma C.15.LetFbe a hypothesis class of feature extractors fromXtoR d with∥F ∥∞ =κ

    The following lemma boundssup f∈F bLwsc(f; bS)− Lwsc(f) . Lemma C.15.LetFbe a hypothesis class of feature extractors fromXtoR d with∥F ∥∞ =κ. Then we have: sup f∈F bLwsc(f; bS)− Lwsc(f) ≤6κ 2d∆(bS)(108) Proof.We expand bLwsc(f; bS)− bLwsc(f) and obtain for anyf∈ F: bLwsc(f; bS)− bLwsc(f) ≤2β E(˜x,˜q),(˜x′,˜q′)∼P(X,Q) 2,x∼A(·|˜x),x′∼A(·|˜x′) h bS(˜x)T :,˜q...

    work page 2021

  5. [7]

    Assume representation dimension d >4r+ 2, Recall λi be the i-th largest eigenvalue of the normalized adjacency matrix. Then, for any ϵ >0 and bf∈ Fsuch that Lwsc( bf)≤ Lwsc(f ∗) +ϵ , we have: ε( bf)≤ε(f ∗) + d ∆2 λ ϵ,(111) where∆ λ ≜λ ⌊3d/4⌋ −λ d is the eigenvalue gap between the⌊3d/4⌋-th and thed-th eigenvalue. D. Related Work Many methods have been deve...

    work page 2024

  6. [8]

    Each runs has been repeated 3 times with different randomly- generated partial labels and we report the mean and std values of last 5 epochs

    Comparisons with each methods on simulated PLL datasets. Each runs has been repeated 3 times with different randomly- generated partial labels and we report the mean and std values of last 5 epochs. Dataset CIFAR-100-H CUB-200 Partial Ratio 0.1 0.5 0.8 0.01 0.05 0.1 PiCO 76.55±0.6874.98±0.4266.38±0.24 74.14±0.2872.12±0.7462.02±1.16 LWS 63.88±0.1059.37±0.2...

    work page

  7. [10]

    Table 10.Ablation studies of our proposed algorithm on CIFAR-100 with different ratio of noisy label and partial label

    The results demonstrate that the proposed WSC loss consistently improves performance, especially under high noise rates, where it increases performance by16.22%and 8.99%, respectively, further validating the effectiveness of the proposed method. Table 10.Ablation studies of our proposed algorithm on CIFAR-100 with different ratio of noisy label and partia...

    work page

  8. [11]

    The second way to construct S is the same as the method used in our main paper

    For the sake of simplicity in this experiment, the first way to construct S is to use a fixed real noise matrix. The second way to construct S is the same as the method used in our main paper. In the third approach, we construct bP(y|x, q) as a convex combination of the original noisy label and the model predictions, with the weights predicted by a two-co...

    work page 2023

Show all 22 references
  1. [13]

    We train our model for 500 epochs and set the initial learning rate be 0.1 and adjust it by Cosine scheduler on both four fine-grained datasets

    Furthermore, for IDPLL settings, we use ResNet-34 pre-trained with ImageNet-1K as our feature extractors following (Yang et al., 2025). We train our model for 500 epochs and set the initial learning rate be 0.1 and adjust it by Cosine scheduler on both four fine-grained datase...

    2025

  2. [14]

    Table 6.Hyper-parameters forsimulated partial label learningused in experiments. Hyper-parameter CIFAR-10 CIFAR-100 (CIFAR-100H) CUB-200 Image Size 32 32 224 Model Wide-ResNet-28-2 Wide-ResNet-28-2 ResNet-18 (ImageNet-1K Pretrained) Batch Size 256 256 256 Learning Rate 0.1 0.1...

    2021

  3. [18]

    They obtain the estimated noise matrix by searching for the matrix that minimizes a specific metric within the matrix family that can linearly represent all noise posteriors

    that has achieved better results. They obtain the estimated noise matrix by searching for the matrix that minimizes a specific metric within the matrix family that can linearly represent all noise posteriors. Specifically, Li et al. 2021 uses volume of estimated noisy matrix a...

    2021

  4. [21]

    The second way to construct S is the same as the method used in our main paper

    For the sake of simplicity in this experiment, the first way to construct S is to use a fixed real partial ratio. The second way to construct S is the same as the method used in our main paper. In the third approach, we construct bP(y|x, q) by using variational label enhanceme...

    2021

  5. [22]

    We report the mean and std values of last 5 epochs.Env.,Sap.denote the environment information and sample information, respectively

    Comparisons with each methods for constructing bS on CIFAR-100 with different ratio of noisy label and partial label. We report the mean and std values of last 5 epochs.Env.,Sap.denote the environment information and sample information, respectively. Type Noisy Label Partial L...

    1977

  6. [92]

    It is direct to see that bLwsc(f;D, bS) is an unbiased estimator of bLwsc(f; bS), to make use of the Rademacher complexity theory to given a generalize bound of this term, we convert the non sum of i.i.d pairwise function to a sum of i.i.d form by using perturbations in U-proc...

    2008

  7. [200]

    Catherine Wah, Steve Branson, Peter Welinder, Pietro Per- ona, and Serge J

    Technical Report CNS-TR-2011- 001, California Institute of Technology, 2011a. Catherine Wah, Steve Branson, Peter Welinder, Pietro Per- ona, and Serge J. Belongie. The caltech-ucsd birds-200- 2011 dataset. 2011b. Haobo Wang, Ruixuan Xiao, Yixuan Li, Lei Feng, Gang Niu, Gang Ch...

    2011

  8. [2013]

    Rep- resentation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Rep- resentation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748,

    arXiv

  9. [2019]

    In recent years, it is the latter (Li et al., 2021; Zhang et al., 2021b; Lin et al.,

    usually fits the noise posterior directly, and then estimates the noise matrix by this estimated noise posterior probability on the data of reliable anchor points. In recent years, it is the latter (Li et al., 2021; Zhang et al., 2021b; Lin et al.,

    2021

  10. [2020]

    Wen et al

    and CC (Feng et al., 2020b) utilize the predictions generated by the predictive model as label information, thus assigning more weights to labels that are more likely to be correct. Wen et al. 2021 propose a family of loss functions for label disambiguation. Wu et al. 2022 per...

    2021

  11. [2021]

    2021 in this section

    For the sake of completeness, we will restate some of the theoretical results in HaoChen et al. 2021 in this section. Lemma C.3.If A is γ-consistent augmentation distribution, and ˆy(x)is corresponding pseudo labeler for augmentation data such that satisfy condition in Definit...

    2021

  12. [2022]

    The former facilitates the formation of well-structured clusters, which in turn enables prototype learning to acquire prototype representations

    integrates contrastive learning with prototype learning. The former facilitates the formation of well-structured clusters, which in turn enables prototype learning to acquire prototype representations. The latter assists in the selection of positive samples for contrastive lea...

    2021

  13. [2023]

    Blaschko, and Andrea Vedaldi

    Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual clas- sification of aircraft.CoRR, abs/1306.5151,

    arXiv

  14. [2024]

    treat the ground-truth labels as latent variables and try to model the entire distribution of all possible labeling entailed by weakly-supervised information, thus allowing a unified solution to deal with NLL and PLL. Although these methods try to uniformly exploit the common ...

    2020

Pith tools

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