Pith. sign in

REVIEW 3 major objections 4 minor 82 references

Fast Rates for Semi-Supervised Learning via Data-Augmentation Graph Regularization

T0 review · 3 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read This paper claims that data augmentation builds a graph on the unlabeled data, and label propagation on that graph achieves expected error C/n_L + R_DA(y), a fast transductive rate rather than the supervised 1/sqrt(n_L) rate.

desk verdict The right connection—fast Johnson–Zhang graph rates placed on the augmentation graph—but the advertised O(1/n_L) rate is not actually derived: the p→∞ step in Remark 17 rests on a false claim about constant diagonals. read the letter →

arxiv 2607.07513 v2 pith:L6EIOIMK submitted 2026-07-08 cs.LG math.STstat.MLstat.TH

Adam M. Oberman This is my paper · ORCID
classification cs.LGmath.STstat.MLstat.TH MSC 68T0568Q3262H30
keywords semi-supervisedlearningdataaugmentationgraphLaplacianregularizationtransductivealgorithmicstabilityfastrateslabelefficiencycut
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

Self-supervised learning famously reaches high accuracy from a small fraction of labels, and this paper tries to say why at the level of finite labeled samples. Its claim: data augmentations define a weighted graph over the unlabeled data, and the downstream classifier is graph-Laplacian-regularized learning on that graph. By carrying a leave-one-out stability argument onto this augmentation graph, the expected transductive error is bounded by $C/n_L$ plus a data-augmentation alignment error $R_{DA}(y)$, the weighted mass of augmentation edges that cross a class boundary. That replaces the supervised $O(1/\sqrt{n_L})$ rate with $O(1/n_L)$ in the number of labels, provided the graph splits into balanced, label-pure components with a non-vanishing margin. A sympathetic reader would take this as a mechanistic explanation of the accuracy-versus-label-count curve: better augmentations shrink $R_{DA}(y)$, so fewer labels are needed.

What carries the argument

The load-bearing object is the augmentation graph: nodes are unlabeled points, edge weights are shared-probability under random augmentation. Its normalized adjacency $W=S^{-1/2}[w_{ij}]S^{-1/2}$ and S-normalized Laplacian $L_S$ define the graph kernel $K^{-1}=\alpha S^{-1}+L_S$. The regularizer $Q(g,S_U)=\sum_k g_k^T K^{-1} g_k$ is the invariance-to-augmentation energy. The argument's engine is a one-point leave-one-out stability lemma (perturbation ≤ $\sigma K_{ii}/(2\lambda n_L)$) which yields a transductive oracle inequality; feeding the label indicator into the oracle term evaluates to the S-normalized cut $R_{DA}(y)$, the channel through which augmentation quality enters the rate.

What would settle it

In a planted-partition graph with known labels and tunable cross-class edge probability, run algorithm (5) over a range of $n_L$ and record the leave-one-out perturbation $|f(x_i)-f^{\setminus i}(x_i)|$ and the 0/1 error. The stability lemma predicts a log-log slope of -1 in $n_L$ and the theorem predicts error ≤ $C/n_L + (\lambda/a)\mathrm{cut}$. A measured slope flatter than -1, or 0/1 error that falls below $(\lambda/a)\mathrm{cut}$ while cut is bounded away from zero, would falsify the fast-rate mechanism.

Watch

Extended reading notes

Core claim

The central result is Theorem 9: for the regularized minimizer with kernel $K^{-1}=\alpha S^{-1}+L_S(G)$ on the augmentation graph, under balanced components and bounded margin, the expected transductive classification error satisfies $E[\mathrm{err}] \leq C/n_L + R_{DA}(y)$, where $R_{DA}(y)=(\lambda/a) \mathrm{cut}(L_S,y)$ is the S-normalized cut of the true labels on the augmentation graph. $R_{DA}(y)$ is zero exactly when augmentations never mix labels, and it grows with the weight of edges joining differently labeled points. The same analysis shows the $1/n_L$ rate is generated by leave-one-out stability: removing one labeled point changes the fitted score by at most $\sigma K_{ii}/(2\lambda n_L)$. In the infinite-data limit with perfect alignment

Load-bearing premise

The fast $1/n_L$ rate rests on Assumption 8: the augmentation graph splits into balanced, label-pure components with a margin between classes that does not vanish; if components are unbalanced, the rate degrades toward $\sqrt{q/n_L}$.

Editorial extensions

If this is right

  • Error splits into a label-budget term C/n_L and an augmentation-quality floor R_DA(y), so once labels outnumber the floor, better augmentations — not more labels — are the lever.
  • R_DA(y) is estimable from a small labeled set and the realized graph before training, making it a selection criterion for ranking augmentation pipelines.
  • If augmentations are perfectly label-consistent (R_DA=0), the error is O(q/n_L), with q the number of balanced components.
  • The streamlined loss is stable and recovers the top-K ideal features; projector, negative-sample, and orthogonality overhead are not needed for the downstream rate.
  • The additive floor is a property of the surrogate loss; the 0/1 classification error follows the floor-free multiplicative form and keeps descending with labels.

Reading between the lines

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

  • The cut estimate could also guide active labeling: since R_DA(y) is a sum over edges crossing label boundaries, labeling a few high-weight boundary nodes and re-estimating the cut may shrink the floor faster than uniform sampling; the paper does not test this.
  • The transductive guarantee would likely extend to fresh points if the empirical augmentation graph converges to the population augmentation operator; then the rate would acquire a graph-approximation term, probably O(1/sqrt(m)) from spectral perturbation theory.
  • If the surrogate floor is real on real data, then a practitioner tuning by surrogate loss may see a plateau caused by augmentation quality even while classification accuracy improves; monitoring the cut would distinguish the two regimes.
  • The bound suggests a prediction for data collection: adding unlabeled points that fall near class boundaries can increase R_DA(y) if augmentations cross those boundaries, so more unlabeled data can sometimes hurt label efficiency.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper studies label propagation on data-augmentation graphs through graph-Laplacian regularization. It claims a fast transductive rate O(1/n_L) in the number of labeled examples, replacing the supervised O(1/sqrt(n_L)), with an expected error bound of the form C/n_L + R_DA(y), where R_DA(y) is a normalized graph cut measuring how much augmentation crosses label boundaries. The main technical steps are a leave-one-out stability bound (Lemma 5), a transductive oracle inequality (Theorem 6), an evaluation of the oracle term as a graph cut (Theorem 9), and a limit result identifying the learned features with the top-K eigenspace of the augmentation kernel (Theorem 12). Experiments include a controlled planted-partition model with known constants and a descriptive CIFAR-10 panel.

Significance. The result, if established, would be significant: it would give a mechanistic, rate-based explanation of the label efficiency of self-supervised learning and make augmentation quality explicit through an estimable graph-cut quantity. The paper is unusually transparent about its debt to Johnson and Zhang (2007), ships code, and clearly separates the synthetic verification from the descriptive CIFAR-10 experiment. However, the advertised fast rate currently rests on an unjustified p-to-infinity step in the appendix, so the central claim is not established by the supplied argument. The contribution is valuable but needs a repaired proof before it can be accepted as stated.

major comments (3)
  1. [Appendix A.5, Remark 17 / Proposition 16(ii)] The transition from the finite-p bound n^{-p/(p+1)} to the advertised O(1/n_L) is not justified. Remark 17 asserts that 'the S-normalization makes the diagonal K_jj constant', but for K=(alpha S^{-1}+L_S)^{-1} this is false: a three-node path with degrees (1,2,1) already yields nonconstant diagonal entries of K. Hence tr_p(K)=(1/m sum_j K_jj^p)^{1/p} generally depends on p, and the p-to-infinity step is invalid without an additional uniform bound on the diagonal of K. Since this p-passage is the only stated mechanism that upgrades Proposition 16(ii) to Theorem 9, the headline O(1/n_L) rate is not derived as written. Please supply a correct proof or add and verify a bounded-diagonal condition.
  2. [Appendix A.4, Lemma 15] The surrogate-to-0/1 conversion is presented as a proof sketch and leaves the 'power-premainder' term untracked. This lemma is load-bearing: it is the only place the margin constant c and Assumption 8(ii) enter the 0/1 error, and it feeds directly into Eq. (15). A complete proof with explicit dependence on a, b, c is needed before the 0/1 statement of Theorem 9 can be verified.
  3. [Assumption 8 / Section 7] The balanced-component condition is load-bearing for the fast rate, but no evidence is given that real augmentation graphs satisfy it. The controlled experiment enforces balance by construction, and the CIFAR-10 panel explicitly makes no floor or cut claim; R_DA(y) is therefore never measured on a realistic SSL pipeline. The conditional nature of the result should be stated more prominently, and at least one empirical check of Assumption 8 on a real augmentation graph would be needed to support the motivating phenomenon.
minor comments (4)
  1. [Throughout] 'CIF AR-10' appears where 'CIFAR-10' is intended.
  2. [Section 2.1] The statement that 'convolution against W realizes the integral operator T_K' is informal; it should be phrased as an empirical-operator approximation with a convergence statement, as used in Theorem 12.
  3. [References] The entry 'Runtian Zhai et al. Contextures' is incomplete; full author list and venue details should be provided.
  4. [Equation (15) / Remark 17] The notation tr_p(K)^p is used before tr_p(K) is defined cleanly; please define it at the first occurrence and distinguish the moment parameter p from the auxiliary parameter p in the optimization.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity: the central bound is a derived, externally sourced inequality, not a fit. Minor circularity signal in Experiment A's by-construction floor verification.

  1. other [Section 7.1 (Experiment A), with Definition 7 and Theorem 9 proof Step 1]
    "Because λ is ours to choose, λ/a is known, and the floor the theorem predicts can be checked against an exact, computed value rather than a fitted one... The squared-loss floor is a line through the origin in the cut (correlation 0.99 across the sweep) with slope below λ/a: the additive prediction λ/a cut of (8), the realized slope falling below λ/a because the oracle does at least as well as the label indicator that furnishes the bound."

    The 'floor the theorem predicts' is R_DA(y)=(λ/a)cut(L_S,y), which Definition 7 defines as the regularizer evaluated at the label indicator. The computed n_L→∞ oracle floor is min_g(loss+λQ); Theorem 9's own Step 1 proves this min is bounded by λ(αs+cut), so the realized slope ≤λ/a is a mathematical consequence of that feasibility step, not an independent empirical test. The paper itself says the oracle 'does at least as well as the label indicator that furnishes the bound.' The experiment verifies only the stability lemma's 1/n slope and the by-construction additive floor; it does not exercise the p→∞ step or pin λ/a on real data.

full rationale

The central derivation is not circular. Lemma 5 and Theorem 6 are self-contained Bregman-divergence/Cauchy-Schwarz stability arguments, and R_DA is a defined graph-cut quantity, not a fitted parameter. The optimized λ in Theorem 9 is asserted to be sample-independent and is not estimated from the CIFAR panel. The fast O(1/n_L) rate is explicitly imported from the external Johnson & Zhang (2007) result ('the fast rate is imported from Johnson and Zhang (2007)...'), and the paper candidly states the contribution is 'not a new rate mechanism but its placement.' The only self-citation (Ghosh et al. 2024) is motivational, not load-bearing for the theorems. The skeptic's p→∞ objection to Remark 17 and the sketch-level Lemma 15 conversion are correctness/proof gaps rather than circularity: a false or incomplete proof step does not make the claimed result equivalent to its input by construction. The paper's Limitations also concede that λ/a is unpinned on real data and that the guarantee is transductive, which further supports a conditional rather than definitional reading. Score 2 reflects only the weak, by-construction character of Experiment A's surrogate-floor verification; no central prediction reduces to a fit or to a self-citation chain.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central theorems lean on an imported published result (Johnson–Zhang 2007) whose key optimization the paper explicitly declines to re-derive. The remaining load-bearing premises are the loss-admissibility assumption (standard) and Assumption 8's balance/margin conditions, asserted without empirical support for real augmentation graphs. The only unpinned numbers are λ, α, and the loss constants a, b, c — the Limitations section concedes λ/a "is not pinned by the theory on real data". No parameters are fitted to data, and R_DA is a defined quantity, not an estimate, in the theory; the Section 8 claim that it can be estimated from a modest labeled sample is a promise without a theorem. No new entities are postulated.

free parameters (3)
  • λ (regularization weight in objective (5)) = chosen by hand; theory's optimized λ* is sample-independent but not pinned on real data
    Appears in the floor R_DA(y) = (λ/a)·cut. Limitations section: "The constant λ/a multiplying the cut is not pinned by the theory on real data", so the quantitative floor is only computable when λ is selected, as in Experiment A.
  • α (ridge coefficient in K^{-1} = αS^{-1} + L_S(G)) = chosen small; value unspecified
    Required for strict positive definiteness of K and folded into the C/n_L term ("α chosen small enough"); no real-data value or selection rule is given.
  • Admissible-loss constants a, b, c (loss scale and margin) = unspecified for actual losses
    Assumption 8(ii) assumes the margin constant c is bounded below by a universal constant; Lemma 15's surrogate-to-0/1 conversion and its "power-premainder" depend on a, b, c, and the paper asserts these hold for standard losses without instantiating them.
assumptions (6)
  • standard math Proposition 16 (Johnson–Zhang 2007 Thms 1/4/5): E[err] ≤ C_p n_L^{-p/(p+1)} (αs+cut)^{p/(p+1)} tr_p(K)^{p/(p+1)} with the stated C_p
    The headline Theorem 9's constant and λ-optimization are explicitly "deferred to Johnson and Zhang (2007)"; the paper declines to reproduce the calculus ("reproving it here would add length without adding insight").
  • domain assumption Assumption 1: σ-admissible loss — convex, σ_ℓ-Lipschitz on a bounded domain
    Used in Lemma 5 and Theorem 6; Appendix A asserts it holds for cosine-similarity, cross-entropy, and regularized least squares without a proof.
  • ad hoc to paper Assumption 8(i): near-zero-cut augmentation graph splits into q pure label-consistent components with m/m1 = Θ(q)
    Upgrades the cut bound from sqrt(q/n_L) to q/n_L; the paper concedes the rate "degrades toward sqrt(q/n_L)" without it. No evidence is given that real augmentation graphs satisfy this.
  • ad hoc to paper Assumption 8(ii): margin constant c of the admissible loss is bounded below uniformly in m
    Keeps the surrogate-to-0/1 conversion constant finite; supported only by Lemma 15's proof sketch, which leaves a "power-premainder" term untracked.
  • domain assumption Assumption 11: spectral gap γ_K = μ_K − μ_{K+1} > 0 for the integral operator T_K of k_DAF
    Required for Theorem 12's Davis–Kahan perturbation step; standard non-degeneracy but unverified for real augmentation kernels.
  • domain assumption R_DA(y) = 0 implies augmentations never mix labels, so the graph splits into pure class components
    Used in Theorem 12 Step 1 to assert the supervised term only selects class-consistent directions and the minimizer's span equals the top-K eigenspace; this is stronger than a small-cut assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fast Rates for Semi-Supervised Learning via Data-Augmentation Graph Regularization." pith.science (2026). https://pith.science/paper/L6EIOIMK

@misc{pith2026260707513,
  author       = {Pith},
  title        = {Pith review of: Fast Rates for Semi-Supervised Learning via Data-Augmentation Graph Regularization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L6EIOIMK}},
  note         = {Machine review of arXiv:2607.07513}
}
abstract

Self-supervised learning matches supervised accuracy from a fraction of the labels, but the labeled-sample efficiency behind this has lacked a theoretical explanation. We provide one. Data augmentation induces a similarity graph on the unlabeled data, so downstream learning on that graph is graph-Laplacian-regularized learning. We prove a fast transductive rate, $O(1/n_L)$ in the number of labels, in place of the supervised $O(1/\sqrt{n_L})$, by carrying the leave-one-out stability apparatus of Johnson and Zhang (JMLR 2007) over to the augmentation graph, and without the unrealistic assumptions of limit-based analyses (exact kernel, generalizing features). The bound makes augmentation quality explicit: the expected error is at most $C/n_L + R_{\mathrm{DA}}(y)$, where the data-augmentation alignment error $R_{\mathrm{DA}}(y)$ is proportional to the graph-cut mass of augmentations that cross a label boundary, so good augmentations let few labels suffice. The analysis uses a streamlined loss that drops the projector, negative-sample, and orthogonality overhead of standard objectives yet still recovers the top-$K$ ideal features in the infinite-data limit, the augmentation-kernel eigenspace studied by Zhai et al. The bound gives a mechanistic account of the accuracy-versus-label-count curve through augmentation quality, verified in a controlled model where the constants are known.

Figures

Figures reproduced from arXiv: 2607.07513 by the authors.

Figure 1
Figure 1. Experiment A (synthetic, known constants; m = 6000, K = 10). Left: the exact nL → ∞ oracle floor against the known cut(LS, y) as the cross-class strength ε is swept. The admissible-surrogate floor is additive, a line through the origin with slope ≤ λ/a (correlation 0.99); the 0/1 error floor stays at zero (the multiplicative form). Centre: the leave-one-out stability |f − f \i | the rate is built from decays at slop… view at source ↗
Figure 2
Figure 2. The phenomenon on CIFAR-10. Transductive accuracy on the unlabeled pool versus the number of labels, for the graph-Laplacian probe (5) and a ridge linear-probe reference, on frozen SimCLR features. Both reach the backbone’s 90.1% ceiling at about 4% of the labels. Descriptive: it establishes the label-efficiency phenomenon the rate explains, with the controlled verification deferred to Experiment A. A Proofs This ap… view at source ↗

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

82 extracted references · 3 linked inside Pith

  1. [1]

    Asymptotic behavior of

    Ahmed El Alaoui , bibsource =. Asymptotic behavior of. Proceedings of the 29th Conference on Learning Theory,. 2016 , bdsk-url-1 =

    2016

  2. [2]

    Generalization Performance of Some Learning Problems in Hilbert Functional Spaces , url =

    Tong Zhang , bibsource =. Generalization Performance of Some Learning Problems in Hilbert Functional Spaces , url =. Advances in Neural Information Processing Systems 14 [Neural Information Processing Systems: Natural and Synthetic,. 2001 , bdsk-url-1 =

    2001

  3. [3]

    Stability and Generalization , url =

    Olivier Bousquet and Andr. Stability and Generalization , url =. J. Mach. Learn. Res. , pages =. 2002 , bdsk-url-1 =

    2002

  4. [4]

    Generalization Analysis of Fredholm Kernel Regularized Classifiers , url =

    Tieliang Gong and Zongben Xu and Hong Chen , bibsource =. Generalization Analysis of Fredholm Kernel Regularized Classifiers , url =. 2017 , bdsk-url-1 =. doi:10.1162/NECO\_A\_00967 , journal =

    doi:10.1162/neco 2017

  5. [5]

    Learning with Fredholm Kernels , url =

    Qichao Que and Mikhail Belkin and Yusu Wang , bibsource =. Learning with Fredholm Kernels , url =. Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada , date-added =. 2014 , bdsk-url-1 =

    2014

  6. [6]

    On the Effectiveness of Laplacian Normalization for Graph Semi-supervised Learning , url =

    Rie Johnson and Tong Zhang , bibsource =. On the Effectiveness of Laplacian Normalization for Graph Semi-supervised Learning , url =. 2007 , bdsk-url-1 =. doi:10.5555/1314498.1314549 , journal =

    arXiv 2007

  7. [7]

    Leave-One-Out Bounds for Kernel Methods , url =

    Tong Zhang , bibsource =. Leave-One-Out Bounds for Kernel Methods , url =. 2003 , bdsk-url-1 =. doi:10.1162/089976603321780326 , journal =

    doi:10.1162/089976603321780326 2003

  8. [8]

    Contrastive and Non-Contrastive Self-Supervised Learning Recover Global and Local Spectral Embedding Methods , url =

    Randall Balestriero and Yann LeCun , bibsource =. Contrastive and Non-Contrastive Self-Supervised Learning Recover Global and Local Spectral Embedding Methods , url =. NeurIPS , date-added =. 2022 , bdsk-url-1 =

    2022

Show all 82 references
  1. [9]

    HaoChen and Colin Wei and Adrien Gaidon and Tengyu Ma , bibsource =

    Jeff Z. HaoChen and Colin Wei and Adrien Gaidon and Tengyu Ma , bibsource =. Provable Guarantees for Self-Supervised Deep Learning with Spectral Contrastive Loss , url =. Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing S...

    2021

  2. [10]

    Johnson and Ayoub El Hanchi and Chris J

    Daniel D. Johnson and Ayoub El Hanchi and Chris J. Maddison , bibsource =. Contrastive Learning Can Find An Optimal Basis For Approximately View-Invariant Functions , url =. The Eleventh International Conference on Learning Representations,. 2023 , bdsk-url-1 =

    2023

  3. [11]

    arXiv preprint arXiv:1806.02215 , title =

    Pfau, David and Petersen, Stig and Agarwal, Ashish and Barrett, David GT and Stachenfeld, Kimberly L , date-added =. arXiv preprint arXiv:1806.02215 , title =

    arXiv

  4. [12]

    A kernel theory of modern data augmentation , year =

    Dao, Tri and Gu, Albert and Ratner, Alexander and Smith, Virginia and De Sa, Chris and R. A kernel theory of modern data augmentation , year =. International conference on machine learning , date-added =

  5. [13]

    Concentration inequalities and moment bounds for sample covariance operators , year =

    Koltchinskii, Vladimir and Lounici, Karim , date-added =. Concentration inequalities and moment bounds for sample covariance operators , year =. Bernoulli , pages =

  6. [14]

    On learning with integral operators

    Rosasco, Lorenzo and Belkin, Mikhail and De Vito, Ernesto , date-added =. On learning with integral operators. , volume =. Journal of Machine Learning Research , number =

  7. [15]

    Functional Data Analysis , date-added =

    Principal components analysis for functional data , url =. Functional Data Analysis , date-added =. 2005 , bdsk-url-1 =. doi:10.1007/0-387-22751-2_8 , isbn =

    2005 doi

  8. [16]

    Linear processes in function spaces: theory and applications , volume =

    Bosq, Denis , date-added =. Linear processes in function spaces: theory and applications , volume =

  9. [17]

    Classes of Linear Operator Theory , volume =

    Gohberg, I , date-added =. Classes of Linear Operator Theory , volume =. Advances and Applications , publisher =

  10. [18]

    Inference for functional data with applications , volume =

    Horv. Inference for functional data with applications , volume =

  11. [19]

    Advances in neural information processing systems , title =

    Bengio, Yoshua and Paiement, Jean-fran. Advances in neural information processing systems , title =

  12. [20]

    Geometric harmonics: a novel tool for multiscale out-of-sample extension of empirical functions , volume =

    Coifman, Ronald R and Lafon, St. Geometric harmonics: a novel tool for multiscale out-of-sample extension of empirical functions , volume =. Applied and Computational Harmonic Analysis , number =

  13. [21]

    Functional analysis , volume =

    Lax, Peter D , date-added =. Functional analysis , volume =

  14. [22]

    An introductory course in functional analysis , year =

    Bowers, Adam and Kalton, Nigel J , date-added =. An introductory course in functional analysis , year =

  15. [23]

    Partial differential equations with numerical methods , volume =

    Larsson, Stig and Thom. Partial differential equations with numerical methods , volume =

  16. [24]

    Neuralef: Deconstructing kernels by deep neural networks , year =

    Deng, Zhijie and Shi, Jiaxin and Zhu, Jun , booktitle =. Neuralef: Deconstructing kernels by deep neural networks , year =

  17. [25]

    arXiv preprint arXiv:2210.12637 , title =

    Deng, Zhijie and Shi, Jiaxin and Zhang, Hao and Cui, Peng and Lu, Cewu and Zhu, Jun , date-added =. arXiv preprint arXiv:2210.12637 , title =

    arXiv

  18. [26]

    An introduction to the theory of reproducing kernel Hilbert spaces , volume =

    Paulsen, Vern I and Raghupathi, Mrinal , date-added =. An introduction to the theory of reproducing kernel Hilbert spaces , volume =

  19. [27]

    Learning similarity with operator-valued large-margin classifiers , volume =

    Maurer, Andreas , date-added =. Learning similarity with operator-valued large-margin classifiers , volume =. The Journal of Machine Learning Research , pages =

  20. [28]

    The elements of statistical learning: data mining, inference, and prediction , volume =

    Hastie, Trevor and Tibshirani, Robert and Friedman, Jerome H and Friedman, Jerome H , date-added =. The elements of statistical learning: data mining, inference, and prediction , volume =

  21. [29]

    Laplacian eigenmaps for dimensionality reduction and data representation , volume =

    Belkin, Mikhail and Niyogi, Partha , date-added =. Laplacian eigenmaps for dimensionality reduction and data representation , volume =. Neural computation , number =

  22. [30]

    Nonlinear component analysis as a kernel eigenvalue problem , volume =

    Sch. Nonlinear component analysis as a kernel eigenvalue problem , volume =. Neural computation , number =

  23. [31]

    Theory of classification: A survey of some recent advances , volume =

    Boucheron, St. Theory of classification: A survey of some recent advances , volume =. ESAIM: probability and statistics , pages =

  24. [32]

    The connection between regularization operators and support vector kernels , volume =

    Smola, Alex J and Sch. The connection between regularization operators and support vector kernels , volume =. Neural networks , number =

  25. [33]

    Oberman , bibsource =

    Ryan Campbell and Chris Finlay and Adam M. Oberman , bibsource =. Adversarial Boot Camp: label free certified robustness in one epoch , url =. CoRR , timestamp =. 2020 , bdsk-url-1 =. 2010.02508 , eprinttype =

    2020 arXiv

  26. [34]

    Razenshteyn and Pengchuan Zhang and Huan Zhang and S

    Hadi Salman and Jerry Li and Ilya P. Razenshteyn and Pengchuan Zhang and Huan Zhang and S. Provably Robust Deep Learning via Adversarially Trained Smoothed Classifiers , url =. Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Proces...

    2019

  27. [35]

    Cohen and Elan Rosenfeld and J

    Jeremy M. Cohen and Elan Rosenfeld and J. Zico Kolter , bibsource =. Certified Adversarial Robustness via Randomized Smoothing , url =. Proceedings of the 36th International Conference on Machine Learning,. 2019 , bdsk-url-1 =

    2019

  28. [36]

    On the mathematical foundations of learning , volume =

    Cucker, Felipe and Smale, Steve , date-added =. On the mathematical foundations of learning , volume =. Bulletin of the American mathematical society , number =

  29. [37]

    Representation learning: A review and new perspectives , volume =

    Bengio, Yoshua and Courville, Aaron and Vincent, Pascal , date-added =. Representation learning: A review and new perspectives , volume =. IEEE transactions on pattern analysis and machine intelligence , number =

  30. [38]

    Advances in neural information processing systems , title =

    Rahimi, Ali and Recht, Benjamin , date-added =. Advances in neural information processing systems , title =

  31. [39]

    Foundations of machine learning , year =

    Mohri, Mehryar and Rostamizadeh, Afshin and Talwalkar, Ameet , date-added =. Foundations of machine learning , year =

  32. [40]

    Partial differential equations: An introduction , year =

    Strauss, Walter A , date-added =. Partial differential equations: An introduction , year =

  33. [41]

    Deep clustering for unsupervised learning of visual features , year =

    Caron, Mathilde and Bojanowski, Piotr and Joulin, Armand and Douze, Matthijs , booktitle =. Deep clustering for unsupervised learning of visual features , year =

  34. [42]

    Support vector machines , year =

    Steinwart, Ingo and Christmann, Andreas , date-added =. Support vector machines , year =

  35. [43]

    A course in functional analysis , volume =

    Conway, John B , date-added =. A course in functional analysis , volume =

  36. [44]

    Advances in neural information processing systems , title =

    Williams, Christopher and Seeger, Matthias , date-added =. Advances in neural information processing systems , title =

  37. [45]

    Advances in neural information processing systems , title =

    Cristianini, Nello and Shawe-Taylor, John and Elisseeff, Andre and Kandola, Jaz , date-added =. Advances in neural information processing systems , title =

  38. [46]

    Learning the kernel matrix with semidefinite programming , volume =

    Lanckriet, Gert RG and Cristianini, Nello and Bartlett, Peter and Ghaoui, Laurent El and Jordan, Michael I , date-added =. Learning the kernel matrix with semidefinite programming , volume =. Journal of Machine learning research , number =

  39. [47]

    Spline models for observational data , year =

    Wahba, Grace , date-added =. Spline models for observational data , year =

  40. [48]

    Learning with kernels: support vector machines, regularization, optimization, and beyond , year =

    Sch. Learning with kernels: support vector machines, regularization, optimization, and beyond , year =

  41. [49]

    Reproducing kernel Hilbert spaces in probability and statistics , year =

    Berlinet, Alain and Thomas-Agnan, Christine , date-added =. Reproducing kernel Hilbert spaces in probability and statistics , year =

  42. [50]

    An introduction to harmonic analysis , year =

    Katznelson, Yitzhak , date-added =. An introduction to harmonic analysis , year =

  43. [51]

    A primer on PDEs: models, methods, simulations , year =

    Salsa, Sandro and Vegni, Federico and Zaretti, Anna and Zunino, Paolo , date-added =. A primer on PDEs: models, methods, simulations , year =

  44. [52]

    Harmonic analysis on semigroups: theory of positive definite and related functions , volume =

    Berg, Christian and Christensen, Jens Peter Reus and Ressel, Paul , date-added =. Harmonic analysis on semigroups: theory of positive definite and related functions , volume =

  45. [53]

    High-dimensional statistics: A non-asymptotic viewpoint , volume =

    Wainwright, Martin J , date-added =. High-dimensional statistics: A non-asymptotic viewpoint , volume =

  46. [54]

    A Simple Framework for Contrastive Learning of Visual Representations , year =

    Ting Chen and Simon Kornblith and Mohammad Norouzi and Geoffrey Hinton , booktitle =. A Simple Framework for Contrastive Learning of Visual Representations , year =

  47. [55]

    Big Self-Supervised Models are Strong Semi-Supervised Learners , year =

    Ting Chen and Simon Kornblith and Kevin Swersky and Mohammad Norouzi and Geoffrey Hinton , booktitle =. Big Self-Supervised Models are Strong Semi-Supervised Learners , year =

  48. [56]

    Olivier J. H. Proceedings of the 37th International Conference on Machine Learning (ICML) , title =. 2020 , note =

    2020

  49. [57]

    Journal of Machine Learning Research , volume =

    Lorenzo Rosasco and Mikhail Belkin and Ernesto De Vito , title =. Journal of Machine Learning Research , volume =

  50. [58]

    Samworth , title =

    Yi Yu and Tengyao Wang and Richard J. Samworth , title =. Biometrika , volume =

  51. [59]

    HaoChen and Colin Wei and Adrien Gaidon and Tengyu Ma , title =

    Jeff Z. HaoChen and Colin Wei and Adrien Gaidon and Tengyu Ma , title =. Advances in Neural Information Processing Systems 34 (NeurIPS) , year =

  52. [60]

    Advances in Neural Information Processing Systems 35 (NeurIPS) , year =

    Randall Balestriero and Yann LeCun , title =. Advances in Neural Information Processing Systems 35 (NeurIPS) , year =

  53. [61]

    Proceedings of the 42nd International Conference on Machine Learning (ICML) , year =

    Runtian Zhai and others , title =. Proceedings of the 42nd International Conference on Machine Learning (ICML) , year =

  54. [62]

    Zico Kolter and Pradeep Ravikumar , title =

    Runtian Zhai and Bingbin Liu and Andrej Risteski and J. Zico Kolter and Pradeep Ravikumar , title =. The Twelfth International Conference on Learning Representations (ICLR) , year =

  55. [63]

    Proceedings of the 36th International Conference on Machine Learning (ICML) , year =

    Sanjeev Arora and Hrishikesh Khandeparkar and Mikhail Khodak and Orestis Plevrakis and Nikunj Saunshi , title =. Proceedings of the 36th International Conference on Machine Learning (ICML) , year =

  56. [64]

    Proceedings of the 40th International Conference on Machine Learning (ICML) , year =

    Yunwen Lei and Tianbao Yang and Yiming Ying and Ding-Xuan Zhou , title =. Proceedings of the 40th International Conference on Machine Learning (ICML) , year =

  57. [65]

    Proceedings of the 36th Conference on Uncertainty in Artificial Intelligence (UAI) , year =

    Jingge Zhu , title =. Proceedings of the 36th Conference on Uncertainty in Artificial Intelligence (UAI) , year =

  58. [66]

    Journal of Machine Learning Research , volume =

    Philippe Rigollet , title =. Journal of Machine Learning Research , volume =

  59. [67]

    Can Semi-Supervised Learning Use All the Data Effectively? A Lower Bound Perspective , booktitle =

    Alexandru Tifrea and Gizem Y. Can Semi-Supervised Learning Use All the Data Effectively? A Lower Bound Perspective , booktitle =. 2023 , note =

    2023

  60. [68]

    Journal of Machine Learning Research , volume =

    Huayi Tang and Yong Liu , title =. Journal of Machine Learning Research , volume =. 2025 , note =

    2025

  61. [69]

    Advances in Neural Information Processing Systems 31 (NeurIPS) , year =

    Vitaly Feldman and Jan Vondrak , title =. Advances in Neural Information Processing Systems 31 (NeurIPS) , year =

  62. [70]

    Proceedings of the 32nd Conference on Learning Theory (COLT) , year =

    Vitaly Feldman and Jan Vondrak , title =. Proceedings of the 32nd Conference on Learning Theory (COLT) , year =

  63. [71]

    Proceedings of the 33rd Conference on Learning Theory (COLT) , year =

    Olivier Bousquet and Yegor Klochkov and Nikita Zhivotovskiy , title =. Proceedings of the 33rd Conference on Learning Theory (COLT) , year =

  64. [72]

    Proceedings of the 17th Annual Conference on Learning Theory (COLT) , year =

    Mikhail Belkin and Irina Matveeva and Partha Niyogi , title =. Proceedings of the 17th Annual Conference on Learning Theory (COLT) , year =

  65. [73]

    Proceedings of the 25th International Conference on Machine Learning (ICML) , year =

    Corinna Cortes and Mehryar Mohri and Dmitry Pechyony and Ashish Rastogi , title =. Proceedings of the 25th International Conference on Machine Learning (ICML) , year =

  66. [74]

    Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD) , year =

    Saurabh Verma and Zhi-Li Zhang , title =. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD) , year =

  67. [75]

    The Tenth International Conference on Learning Representations (ICLR) , year =

    Yifei Wang and Qi Zhang and Yisen Wang and Jiansheng Yang and Zhouchen Lin , title =. The Tenth International Conference on Learning Representations (ICLR) , year =

  68. [76]

    Dhillon and Sujay Sanghavi and Qi Lei , title =

    Shuo Yang and Yijun Dong and Rachel Ward and Inderjit S. Dhillon and Sujay Sanghavi and Qi Lei , title =. Proceedings of the 26th International Conference on Artificial Intelligence and Statistics (AISTATS) , year =

  69. [77]

    The Eleventh International Conference on Learning Representations (ICLR) , year =

    Rattana Pukdee and Dylan Sam and Maria-Florina Balcan and Pradeep Ravikumar , title =. The Eleventh International Conference on Learning Representations (ICLR) , year =

  70. [78]

    Lafferty , title =

    Xiaojin Zhu and Zoubin Ghahramani and John D. Lafferty , title =. Proceedings of the 20th International Conference on Machine Learning (ICML) , year =

  71. [79]

    Journal of Machine Learning Research , volume =

    Mikhail Belkin and Partha Niyogi and Vikas Sindhwani , title =. Journal of Machine Learning Research , volume =

  72. [80]

    Advances in Neural Information Processing Systems 36 (NeurIPS) , year =

    Yijun Dong and Kevin Miller and Qi Lei and Rachel Ward , title =. Advances in Neural Information Processing Systems 36 (NeurIPS) , year =

  73. [81]

    Proceedings of the 42nd International Conference on Machine Learning (ICML) , year =

    Jun Chen and Hong Chen and Yonghua Yu and Yiming Ying , title =. Proceedings of the 42nd International Conference on Machine Learning (ICML) , year =

  74. [82]

    Oberman and Blake A

    Arna Ghosh and Kumar Krishna Agrawal and Shagun Sodhani and Adam M. Oberman and Blake A. Richards , title =. Advances in Neural Information Processing Systems 37 (NeurIPS) , year =

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.