Pith. sign in

REVIEW 4 major objections 5 minor 35 references

JojoSCL: Shrinkage Contrastive Learning for single-cell RNA sequence Clustering

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

Pith's one-line read JojoSCL claims best single-cell clustering scores on nine of ten datasets by adding a shrinkage loss to contrastive learning.

desk verdict The paper's SURE derivation is algebraically wrong and the MSE guarantee compares to the fitted mean, so the theory is unsound; the empirical recipe is plausible but the '9/10' claim is overstated. read the letter →

arxiv 2506.00410 v1 pith:IUW6OU4Q submitted 2025-05-31 cs.LG q-bio.GNstat.ML

classification cs.LGq-bio.GNstat.ML MSC 62H3062F1568T07
keywords single-cellRNAsequencingclusteringcontrastivelearningshrinkageestimatorStein'sunbiasedriskestimatehierarchicalBayesianmodelself-supervisedcelltypeidentification
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

JojoSCL sets out to show that a shrinkage estimator borrowed from hierarchical Bayesian statistics can be grafted onto contrastive learning for single-cell RNA sequencing (scRNA-seq) clustering. The paper claims that penalizing embeddings with a Stein's Unbiased Risk Estimate (SURE) loss, which pulls each cell toward its cluster centroid while accounting for uncertainty in that centroid, reduces intra-cluster dispersion and sharpens both instance-level and cluster-level contrastive losses. Across ten scRNA-seq datasets, the resulting model reports the best Adjusted Rand Index and Normalized Mutual Information on nine of them, including gains over graph-based, deep clustering, and earlier contrastive baselines. If the claim holds, it would mean a single principled variance-shrinkage term, not a larger model, is enough to push state-of-the-art cell-type clustering further.

What carries the argument

The load-bearing mechanism is the SURE loss $L_{\mathrm{SURE}}$ built from the hierarchical MAP shrinkage estimator $\hat{\theta}_{\mathrm{MAP}}$ and Stein's unbiased risk formula. The paper computes the SURE expression as $\frac{\sigma_k^2}{\tau_k^2+\sigma_k^2}(\|\mu_k - h_{i,k}^a\|_2^2 + P(\tau_k^2 - \sigma_k^2))$ for each cluster, estimates $\hat{\mu}_k$ as the cluster mean and $\hat{\tau}_k^2 = \hat{\sigma}_k^2 / N_k$ from the temporarily assigned features, and uses the aggregate over samples as a loss. Its role is to concentrate the learned embeddings around their cluster centroids, which the paper argues improves cosine-similarity separation between positive and negative pairs in both instance-level and cluster-level contrastive learning.

What would settle it

Compute, on a held-out set, the true mean squared error of the MAP-shrunken embeddings and compare it with the SURE value reported by the paper's formula across many random data draws; if the two diverge systematically, the SURE term is not an unbiased risk estimate. A simpler comparative test: replace the estimated $\hat{\mu}_k$ and $\hat{\tau}_k^2$ with values from an independent batch of cells and see whether the clustering gain persists.

Watch

Extended reading notes

Core claim

The central discovery is that including a SURE-optimized hierarchical shrinkage term in the training objective makes contrastive embeddings more separable. The paper derives the MAP estimator $\hat{\theta}_{\mathrm{MAP}}(\sigma, X; \mu, \tau) = \frac{\tau^2}{\tau^2+\sigma^2}X + \frac{\sigma^2}{\tau^2+\sigma^2}\mu$ for a normal likelihood with a normal prior, and uses SURE to get an unbiased estimate of its squared-error risk. Applied to K clusters, with K-means providing temporary labels and with the prior mean and variance estimated from the same batch of features, the resulting $L_{\mathrm{SURE}}$ penalizes each sample's squared distance to its cluster centroid in a variance-weighted way. This loss is added to the instance-level and cluster-level contrastive losses, and the paper reports that this combination reaches the best ARI/NMI on nine of ten datasets, with average ARI 0.8300 and NMI 0.8207 across all ten.

Load-bearing premise

The load-bearing premise is that the SURE formula remains an unbiased risk estimate when the prior mean and variance in the shrinkage term are estimated from the same feature vectors being shrunk; if that circularity breaks, the claimed shrinkage guarantee is not supported.

Editorial extensions

If this is right

  • Adding $L_{\mathrm{SURE}}$ widens the mean cosine-similarity gap between positive and negative pairs in 9 of 10 datasets, with a paired t-test p-value of 0.0061.
  • The full three-term loss beats any two-term combination in 9 of 10 datasets, and $L_{\mathrm{SURE}}$ combined with the instance loss outperforms the cluster loss alone in ablation comparisons.
  • The model retains higher NMI than baselines under random and stratified downsampling of the Klein dataset, including at 80% data removal.
  • Gaussian noise and masking during augmentation helps NMI on 7 of 10 datasets, particularly large-subtype and small-cell-count datasets.
  • Convergence to stable clustering is reached within a few hundred epochs on the datasets examined.

Reading between the lines

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

  • If the SURE derivation were sound, the same shrinkage loss could be transferred to other sparse high-dimensional biological assays such as ATAC-seq or methylation arrays, where cluster centroids are similarly noisy.
  • The method implicitly tests a broader hypothesis: explicit variance reduction of embeddings may matter more than adding more negative samples or stronger augmentations in contrastive learning.
  • A natural extension would replace the hard K-means temporary labels in the SURE term with soft or learned assignments, which could smooth the centroid estimates and possibly remove the circularity in estimating prior parameters.
  • The reported paired t-tests on pairwise-similarity gaps suggest that $L_{\mathrm{SURE}}$ acts mainly by increasing separation margins; an independent check on held-out risk could show whether the benefit is truly due to unbiased risk minimization.
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. The paper proposes JojoSCL, a self-supervised contrastive clustering framework for scRNA-seq data. The method combines instance-level and cluster-level contrastive losses with an additional shrinkage loss intended to pull feature embeddings toward cluster centroids. The shrinkage estimator is motivated by hierarchical Bayesian modeling and is claimed to be optimized via Stein's Unbiased Risk Estimate (SURE). The authors report state-of-the-art ARI and NMI results on ten scRNA-seq datasets, together with robustness and ablation studies, and they release code at a public repository.

Significance. If the SURE-based shrinkage claim were valid, the paper would make a useful conceptual contribution: it would connect hierarchical Bayesian shrinkage to contrastive representation learning for a high-dimensional, noisy data type. The empirical study is broad, spans ten datasets, includes ablations and robustness analyses, and the code is made publicly available. However, the theoretical foundation is not valid as written: the central MSE and SURE derivations contain algebraic and statistical errors, and the headline empirical claim is contradicted by the paper's own table. The contribution therefore reduces to an ad hoc variance-penalty heuristic whose empirical gains, while suggestive, are not supported by the claimed theory. The strengths are the breadth of experiments and the availability of code, not the theoretical apparatus.

major comments (4)
  1. [II.B, Eq. (10)] The MSE comparison in Eq. (10) is not a comparison of risks with respect to the true parameter θ. The displayed quantity is E[||(τ²/(τ²+σ²))(X-μ)||²], i.e., the distance to the prior mean μ, whereas both θMLE = X and the MAP estimator should be evaluated against θ. Since the MAP estimator is constructed by shrinking toward μ, this comparison is tautological and cannot establish a reduction in mean squared error for estimating θ. A correct calculation would need to evaluate E[||aX+bμ-θ||²] for a=τ²/(τ²+σ²) and b=σ²/(τ²+σ²), which is not what Eq. (10) reports.
  2. [II.B, Eq. (12)] The SURE simplification in Eq. (12) is algebraically incorrect. Substituting θMAP = aX+bμ into the SURE expression of Eq. (11) gives -Pσ² + (σ²/(τ²+σ²))² ||μ-X||² + 2Pσ²τ²/(τ²+σ²), so the coefficient of ||μ-X||² should be σ⁴/(τ²+σ²)². Equation (12) instead has σ²/(τ²+σ²); the two expressions coincide only when σ² = τ²+σ², which is generally false. The subsequent SURE loss in Eqs. (15) and (20)-(21) inherits this error.
  3. [III.B, Eqs. (16)-(17)] The derivation treats the prior parameters as constants in the SURE divergence term, but in the actual algorithm μk and τk are estimated from the same features ha_i being shrunk. Equation (16) sets μk to the intra-cluster sample mean, and Eq. (17) invokes the CLT to set τ²k = σ²k/Nk, but no sampling model or independence condition is stated, and the plug-in makes the prior data-dependent. Consequently, the derivative ∂μ/∂Xp is nonzero and is omitted from Eq. (12), so L_SURE is not an unbiased estimate of the risk of the implemented shrinkage estimator. This is load-bearing because the paper's central claim is that the shrinkage estimate is 'optimized using SURE.'
  4. [IV.B, Table II] The statement that 'JojoSCL consistently outperforms all other methods in ARI and NMI across 9 out of 10 datasets' is not supported by Table II. JojoSCL achieves the best ARI on 8 datasets (Mouse and Yan are won by baselines) and the best NMI on 7 datasets (Macosko, Mouse, and Yan are won by baselines). For Macosko, scziDesk reports a higher NMI (0.8247) than JojoSCL (0.8145). The empirical summary and the derived claims about average performance gains should be corrected to match the table.
minor comments (5)
  1. [III.B, Eq. (18)] The variance estimator in Eq. (18) sums over p=1,...,P inside a term indexed by p and labeled σ²p,k; the summation should presumably be over the Nk samples in the cluster, not over the P feature dimensions. Please correct the indices and the estimator definition.
  2. [IV.A and IV.B] Table II reports results 'based on 10 consecutive runs' but no standard deviations or statistical significance tests are reported for the main comparisons, despite later paired t-tests being used in the robustness analyses.
  3. [IV.B, bullet 4 and Fig. 2] The claim of 'significantly faster convergence and reduced overall training time' is not supported by evidence: Fig. 2 shows only JojoSCL's own convergence curves, with no comparison to competing methods or wall-clock timings.
  4. [IV.D and Figs. 3-4] The text states that under downsampling 'JojoSCL's NMI remains higher than competing methods across all levels,' but the figures do not show competing methods, so this claim is not demonstrated.
  5. [IV.A and III.E] The hyperparameters α, β, τI, τC, and m appear in the loss definitions but their values or tuning procedure are not reported in the experimental setup, which reduces reproducibility.

Circularity Check

2 steps flagged · score 7.0 of 10

Central shrinkage guarantees reduce to definitions: Eq. (10) measures MSE against the shrinkage target, and L_SURE is a within-cluster-distance penalty evaluated on the same fitted centroids; the SURE derivation is also algebraically unsound.

  1. self definitional [Section II.B, Eq. (10)]
    "MSE(θMLE) = E[∥X − µ∥2 2], MSE(θMAP) = E[∥ τ2 τ2+σ2 X + σ2 τ2+σ2 µ − µ∥2 2] = E[∥ τ2 τ2+σ2 (X − µ)∥2 2] (10) where τ2/(τ2+σ2) ≤ 1. This result supports the assertion that the MAP estimator ... achieves a guaranteed lower mean squared error (MSE) compared to the MLE estimator."

    The quantity labeled MSE is the squared distance to the prior mean µ, not to the true θ. Since θMAP is exactly an interpolation between X and µ, its distance to µ is the fraction τ²/(τ²+σ²) of X−µ; the claimed 'guaranteed lower MSE' follows immediately from the definition of the target and does not establish that θMAP estimates θ better. The shrinkage benefit invoked throughout the paper therefore rests on an identity rather than on a statistical risk comparison.

  2. fitted input called prediction [Section III.B, Eq. (21), and Section IV.E]
    "LSURE = Σ_i Σ_k 1{i∈k} [ σ̂²_k/(σ̂²_k/N_k + σ̂²_k) (∥ h^a_{i,k} − h̄^a_{i,k} ∥²_2 + P(σ̂²_k/N_k − σ̂²_k)) ] (21) ... The loss function LSURE imposes a penalty based on the aggregate dispersion of multi-centroids in clustering tasks."

    Here µ̂_k is computed in Eq. (16) as the mean of the same h^a_{i,k} features being shrunk, and the indicator assigns points using K-means on those same features. The SURE loss is therefore a weighted within-cluster sum of squared distances to the fitted centroids. Minimizing it must reduce intra-cluster dispersion, so the reported benefit of L_SURE (Table IV, 'increases the mean of the difference in cosine similarity ... in 9 out of 10 datasets') partially confirms the loss's own objective rather than an independent consequence of shrinkage. The fitted centroid and the evaluated dispersion are the same quantity, making the tighter-cluster prediction forced by construction.

full rationale

The paper's benchmark comparison in Table II is externally grounded: ARI/NMI values are computed against independent baselines, so the empirical clustering results are not circular. The claim '9 out of 10 datasets' is contradicted by Table II, where JojoSCL is best in 8 datasets for both ARI and NMI, but that is a factual error rather than a circularity. However, the two theoretical pillars that motivate the method are circular or self-confirming. Equation (10) defines the 'MSE' of the MLE as E||X−µ||², i.e. the distance to the same prior mean µ toward which θMAP shrinks; the guaranteed reduction is then the trivial inequality τ²/(τ²+σ²)<1, not a result about estimating the true θ. Likewise, Eqs. (15)-(21) construct L_SURE from the within-cluster squared distance to µ̂_k, where µ̂_k is the mean of the same h_i features and cluster membership comes from K-means on those features; minimizing L_SURE necessarily reduces intra-cluster dispersion, so the ablation in Section IV.E partly verifies the loss's own definition. Separately, Eq. (12) does not follow from substituting Eq. (7) into Eq. (11): the coefficient of ||µ−X||² should be σ⁴/(τ²+σ²)², not σ²/(τ²+σ²), and Eqs. (16)-(17) estimate µ and τ from the same features whose derivative is omitted. Those defects are correctness issues rather than circularity, but they eliminate the independent first-principles support for the SURE label. Overall, the central shrinkage guarantee reduces to construction, while the benchmark numbers remain independent, giving a partial-circularity score of 7.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities beyond a new loss term. The central assumptions are the normality model for scRNA-seq features, the empirical Bayes prior, and the specific identifications of τ² and the SURE divergence term, all of which are used to turn the loss into a computable regularizer.

free parameters (4)
  • α (instance-level loss weight) = not reported
    Introduced in the final loss Eq (31); no value or search range is given in the text.
  • β (cluster-level loss weight) = not reported
    Same as α; controls the balance between instance-level and cluster-level contrastive losses.
  • temperature parameters τ_I and τ_C = not reported
    Used in the contrastive losses (23) and (28); their values are not stated.
  • momentum coefficient m = not reported
    Used in the momentum encoder update Eq (13); the value is not stated.
assumptions (5)
  • domain assumption scRNA-seq features within a cluster are i.i.d. normal with a common variance and a consistent covariance matrix across clusters
    Assumed at the start of Section II. Single-cell count data are not normally distributed, so this is a strong simplification.
  • domain assumption Cluster centroids follow a normal prior with mean μ_k and variance τ²_k, with μ_k and τ²_k estimated from the same data
    Eqs (4), (5), and (14). Empirical Bayes estimation is standard, but here the estimates are not independent of the features used to evaluate SURE.
  • ad hoc to paper τ²_k = σ²_k / N_k, equating the prior variance of the centroid with the sampling variance of its sample mean
    Eq (17). No derivation is provided, and the identity does not follow from any stated prior model. It is introduced to make the SURE loss computable.
  • ad hoc to paper The SURE derivative treats μ_k as a constant even though μ_k is the mean of the same features
    In Eq (12), the divergence term differentiates only the explicit X, not μ, because μ is a function of the same ha_i. This omission is necessary for the simplified SURE expression but invalidates the unbiasedness claim.
  • domain assumption K-means temporal labels used in Eq (21) are the correct cluster assignments for computing the SURE loss
    Section III.B. If the temporal labels are noisy, the loss pulls embeddings toward wrong centroids during training, and the shrinkage interpretation breaks down.

how reviews work

0 comments
Cite this review

Pith. "Pith review of JojoSCL: Shrinkage Contrastive Learning for single-cell RNA sequence Clustering." pith.science (2026). https://pith.science/paper/IUW6OU4Q

@misc{pith2026250600410,
  author       = {Pith},
  title        = {Pith review of: JojoSCL: Shrinkage Contrastive Learning for single-cell RNA sequence Clustering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IUW6OU4Q}},
  note         = {Machine review of arXiv:2506.00410}
}
read the original abstract

Single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of cellular processes by enabling gene expression analysis at the individual cell level. Clustering allows for the identification of cell types and the further discovery of intrinsic patterns in single-cell data. However, the high dimensionality and sparsity of scRNA-seq data continue to challenge existing clustering models. In this paper, we introduce JojoSCL, a novel self-supervised contrastive learning framework for scRNA-seq clustering. By incorporating a shrinkage estimator based on hierarchical Bayesian estimation, which adjusts gene expression estimates towards more reliable cluster centroids to reduce intra-cluster dispersion, and optimized using Stein's Unbiased Risk Estimate (SURE), JojoSCL refines both instance-level and cluster-level contrastive learning. Experiments on ten scRNA-seq datasets substantiate that JojoSCL consistently outperforms prevalent clustering methods, with further validation of its practicality through robustness analysis and ablation studies. JojoSCL's code is available at: https://github.com/ziwenwang28/JojoSCL.

Figures

Figures reproduced from arXiv: 2506.00410 by the authors.

Figure 1
Figure 1. The overview of the proposed model. The model consists of data augmentation and shrinkage contrastive learning. Our approach utilizes a shrinkage [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The convergence of NMI and loss across 400 epochs for JojoSCL [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Clustering results for the Klein dataset using JojoSCL with random [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Clustering results for the Klein dataset using JojoSCL with stratified [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The separation between positive and negative pairs and the growing [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 32 canonical work pages

  1. [1]

    Applications of single-cell RNA sequencing in drug discovery and development,

    B. Van de Sande, J. S. Lee, E. Mutasa-Gottgens, et al., “Applications of single-cell RNA sequencing in drug discovery and development,” Nat. Rev. Drug Discov. , vol. 22, pp. 496–520, 2023

  2. [2]

    Cryo-electron tomog- raphy on focused ion beam lamellae transforms structural cell biology,

    C. Berger, N. Premaraj, R. B. G. Ravelli, et al., “Cryo-electron tomog- raphy on focused ion beam lamellae transforms structural cell biology,” Nat. Methods , vol. 20, pp. 499–511, 2023

  3. [3]

    Best practices for single- cell analysis across modalities,

    L. Heumos, A. C. Schaar, C. Lance, et al., “Best practices for single- cell analysis across modalities,” Nat. Rev. Genet. , vol. 24, pp. 550–572, 2023

  4. [4]

    CIDR: Ultrafast and accurate clustering through imputation for single-cell RNA-seq data,

    P. Lin, M. Troup, and J. W. Ho, “CIDR: Ultrafast and accurate clustering through imputation for single-cell RNA-seq data,” Genome Biology, vol. 18, no. 1, pp. 1–11, 2017

  5. [5]

    Vi- sualization and analysis of single-cell RNA-seq data by kernel-based similarity learning,

    B. Wang, J. Zhu, E. Pierson, D. Ramazzotti, and S. Batzoglou, “Vi- sualization and analysis of single-cell RNA-seq data by kernel-based similarity learning,” Nature Methods, vol. 14, no. 4, pp. 414–416, April 2017

  6. [6]

    SAFE- clustering: Single-cell aggregated (from ensemble) clustering for single- cell RNA-seq data,

    Y . Yang, R. Huh, H. W. Culpepper, Y . Lin, M. I. Love, and Y . Li, “SAFE- clustering: Single-cell aggregated (from ensemble) clustering for single- cell RNA-seq data,” Bioinformatics, vol. 35, pp. 1269–1277, February 2018

  7. [7]

    Spatial reconstruction of single-cell gene expression data,

    R. Satija, J. A. Farrell, D. Gennert, A. F. Schier, and A. Regev, “Spatial reconstruction of single-cell gene expression data,” Nature Biotechnology, vol. 33, no. 5, pp. 495–502, May 2015

  8. [8]

    Deep learning,

    Y . LeCun, Y . Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, May 2015

Show all 35 references
  1. [9]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778

  2. [10]

    Unsupervised deep embedding for clustering analysis,

    J. Xie, R. Girshick, and A. Farhadi, “Unsupervised deep embedding for clustering analysis,” in Proceedings of the International Conference on Machine Learning (ICML), 2016, pp. 478–487

  3. [11]

    DeepImpute: An accurate, fast, and scalable deep neural network method to impute single-cell RNA-seq data,

    C. Arisdakessian, O. Poirion, B. Yunits, X. Zhu, and L. X. Garmire, “DeepImpute: An accurate, fast, and scalable deep neural network method to impute single-cell RNA-seq data,” Genome Biology, vol. 20, no. 1, pp. 211, October 2019

  4. [12]

    Single-cell RNA-seq denoising using a deep count autoencoder,

    G. Eraslan, L. M. Simon, M. Mircea, N. S. Mueller, and F. J. Theis, “Single-cell RNA-seq denoising using a deep count autoencoder,” Nature Communications, vol. 10, no. 1, pp. 1–14, 2019

  5. [13]

    Clustering single-cell RNA- seq data with a model-based deep learning approach,

    T. Tian, J. Wan, Q. Song, and Z. Wei, “Clustering single-cell RNA- seq data with a model-based deep learning approach,” Nature Machine Intelligence, vol. 1, no. 4, pp. 191–198, April 2019

  6. [14]

    Contrastive clustering,

    Y . Li, P. Hu, Z. Liu, D. Peng, J. T. Zhou, and X. Peng, “Contrastive clustering,” in Proceedings of the AAAI Conference on Artificial Intel- ligence, 2021, pp. 8547–8555

  7. [15]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” in Proceedings of the International Conference on Machine Learning (ICML), 2020, pp. 1597– 1607

  8. [16]

    Dense contrastive learning for self-supervised visual pretraining,

    X. Wang, R. Zhang, C. Shen, T. Kong, and L. Li, “Dense contrastive learning for self-supervised visual pretraining,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2021, pp. 3024–3033

  9. [17]

    Contrastive self-supervised clustering of scRNA-seq data,

    M. Ciortan and M. Defrance, “Contrastive self-supervised clustering of scRNA-seq data,” BMC Bioinformatics, vol. 22, no. 1, pp. 280, 2021

  10. [18]

    scNAME: Neighborhood contrastive clustering with ancillary mask estimation for scRNA-seq data,

    H. Wan, L. Chen, and M. Deng, “scNAME: Neighborhood contrastive clustering with ancillary mask estimation for scRNA-seq data,” Bioin- formatics, vol. 38, no. 6, pp. 1575–1583, March 2022

  11. [19]

    Self-supervised contrastive learning for integrative single cell RNA-seq data analysis,

    W. Han, Y . Cheng, J. Chen, H. Zhong, Z. Hu, S. Chen, L. Zong, L. Hong, T.-F. Chan, I. King, et al., “Self-supervised contrastive learning for integrative single cell RNA-seq data analysis,” Briefings in Bioin- formatics, vol. 23, no. 5, pp. bbac377, 2022

  12. [20]

    Representation learning with contrastive predictive coding,

    A. van den Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,” arXiv preprint arXiv:1807.03748, July 2018

  13. [21]

    Validation of noise models for single-cell transcriptomics,

    D. Gr ¨un, L. Kester, and A. van Oudenaarden, “Validation of noise models for single-cell transcriptomics,” Nature Methods, vol. 11, pp. 637–640, June 2014

  14. [22]

    SCCCL: Single-cell data clustering based on self-supervised contrastive learning,

    L. Du, R. Han, B. Liu, Y . Wang, and J. Li, “SCCCL: Single-cell data clustering based on self-supervised contrastive learning,” IEEE/ACM Transactions on Computational Biology and Bioinformatics, vol. 20, no. 03, pp. 2233–2241, 2023

  15. [23]

    Momentum contrast for unsupervised visual representation learning,

    K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum contrast for unsupervised visual representation learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 9729–9738

  16. [24]

    Group equivariant convolutional networks,

    T. Cohen and M. Welling, “Group equivariant convolutional networks,” in Proceedings of the International Conference on Machine Learning (ICML), 2016, pp. 2990–2999

  17. [25]

    Choice of hierarchical priors: Admissibility in estimation of normal means,

    J. O. Berger, W. E. Strawderman, et al., “Choice of hierarchical priors: Admissibility in estimation of normal means,” The Annals of Statistics, vol. 24, no. 3, pp. 931–951, 1996

  18. [26]

    Estimation of the mean of a multivariate normal distribu- tion,

    C. M. Stein, “Estimation of the mean of a multivariate normal distribu- tion,” The Annals of Statistics, pp. 1135–1151, 1981

  19. [27]

    C-SURE: Shrinkage estimator and prototype classifier for complex-valued deep learning,

    R. Chakraborty, Y . Xing, M. Duan, and S. X. Yu, “C-SURE: Shrinkage estimator and prototype classifier for complex-valued deep learning,” in Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 360–367, 2020

  20. [28]

    Can contrastive learning avoid shortcut solutions?

    L. Sun, J. Robinson, K. Yu, K. Batmanghelich, S. Jegelka, and S. Sra, “Can contrastive learning avoid shortcut solutions?” Advances in Neural Information Processing Systems, vol. 34, pp. 4974–4986, 2021

  21. [29]

    Rethinking weak supervision in helping contrastive learning,

    J. Cui, W. Huang, Y . Wang, and Y . Wang, “Rethinking weak supervision in helping contrastive learning,” in Proceedings of the 40th International Conference on Machine Learning , vol. 202, pp. 6448–6467, Jul. 2023

  22. [30]

    Do more negative samples nec- essarily hurt in contrastive learning?

    P. Awasthi, N. Dikkala, and P. Kamath, “Do more negative samples nec- essarily hurt in contrastive learning?” in Proceedings of the International Conference on Machine Learning, 2022

  23. [31]

    Less can be more in con- trastive learning,

    J. Mitrovic, M. Rey, and B. McWilliams, “Less can be more in con- trastive learning,” in Proceedings on “I Can’t Believe It’s Not Better!” at NeurIPS Workshops, vol. 137, pp. 70–75, 2020

  24. [32]

    Estimation with quadratic loss,

    W. James and C. Stein, “Estimation with quadratic loss,” in Break- throughs in Statistics, pp. 443–460, Springer, 1992

  25. [33]

    Stein’s estimation rule and its competitors—an empirical Bayes approach,

    B. Efron and C. Morris, “Stein’s estimation rule and its competitors—an empirical Bayes approach,” Journal of the American Statistical Associ- ation, vol. 68, no. 341, pp. 117–130, 1973

  26. [34]

    Exact Bayesian inference on discrete models via probability generating functions: A probabilistic programming approach,

    F. Zaiser, A. Murawski, and C.-H. L. Ong, “Exact Bayesian inference on discrete models via probability generating functions: A probabilistic programming approach,” in Advances in Neural Information Processing Systems, vol. 36, pp. 2427–2462, 2023

  27. [35]

    Deep soft K-means clustering with self-training for single-cell RNA sequence data,

    L. Chen, W. Wang, Y . Zhai, and M. Deng, “Deep soft K-means clustering with self-training for single-cell RNA sequence data,” NAR Genomics and Bioinformatics, vol. 2, no. 2, pp. lqaa039, 2020

Pith tools

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