Pith. sign in

REVIEW 5 major objections 6 minor 15 references

A kernel-based SSL framework scales to million-sample datasets and makes each embedding traceable to specific training examples.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 14:32 UTC pith:KCMKHLUK

load-bearing objection Reasonable kernel-SSL engineering, but the headline CIFAR-10 comparison is confounded by an ImageNet-pretrained eNTK and the promised bias audit is missing. the 5 major comments →

arxiv 2509.24467 v4 pith:KCMKHLUK submitted 2025-09-29 cs.LG stat.ML

Interpretable Self-Supervised Learning via Representer Landmarks and Nystr\"om Approximation

Maedeh Zarvandi , Michael Timothy , Theresa Wasserer , Debarghya Ghoshdastidar This is my paper
classification cs.LG stat.ML
keywords self-supervised learningkernel methodsNyström approximationrepresenter theoreminterpretabilityneural tangent kernelrepresentation learninginfluence scores
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

KREPES is a proposed framework for doing self-supervised representation learning with kernel machines instead of deep networks. The paper argues that by restricting the model to a Nyström subset of 'representer landmarks' and training the coefficient matrix with gradient descent, standard SSL objectives (SimCLR, BYOL, VICReg, Barlow Twins, and others) can be optimized at up to one-million-sample scale. Because the representer theorem makes the learned embedding a weighted sum of kernel similarities to the landmarks, interpretation is not an add-on: the trained weights tell you which unlabeled examples drive a given representation, and the paper introduces influence scores and concept-alignment profiles to read them out. Empirically the paper reports accuracy close to or better than the corresponding neural networks on CIFAR-10, Adult, Higgs, and CoverType, with an order of magnitude fewer trainable parameters. If correct, this would give large-scale SSL a route to transparency, including auditing for demographic bias in unlabeled data.

Core claim

The central discovery is that a kernel-based representation learner can be made practical and interpretable at the same time: for any of a wide class of SSL losses, optimizing over the finite-dimensional space spanned by m landmark features yields solutions f(x)=Σ_l α_l k(c_l,x)+γ, and the coefficient matrix à learned by gradient descent is itself the explanation. The Nyström approximation (m≈2000 landmarks), a principal-component initialization Ã0=U_hΛ_h^{-1/2}, and generalized Gauss–Newton preconditioners make training tractable; the representer theorem guarantees the optimal map lies in the landmark span, so influence scores (local kernel similarity times global column norm) and concept-c

What carries the argument

The representer theorem for kernels in the Nyström subspace: any optimal embedding has the form f(x)=Σ_{l=1}^{m} α_l k(c_l, x)+γ, where c_l are the selected landmarks (a subset of training points) and α_l are learned row vectors. This identity does double duty: it reduces an infinite-dimensional search to the m×h matrix Ã, making gradient-based SSL tractable, and it gives every coordinate of the embedding a direct attribution to specific training examples. The accompanying Nyström low-rank factorization K≈K_nm K_mm^{-1}K_mn and the principal-component initialization Ã0=U_hΛ_h^{-1/2} are what let the method reach 1M samples; the GGN preconditioners are what let it optimize arbitrary SSL losse

Load-bearing premise

The accuracy numbers depend on the kernel being built from a network that is already pretrained on the same type of data (for images, a ResNet34 trained on ImageNet), and the paper never tests a randomly initialized kernel, so the parity with neural networks may be coming from the pretrained features rather than from the kernel self-supervised machinery.

What would settle it

Train KREPES on CIFAR-10 with an empirical NTK computed from a randomly initialized ResNet34 (no ImageNet pretraining), keeping all other settings identical; if the 88–90% downstream accuracy collapses toward the 41–60% of the neural-network baseline, the paper's scaling claim is carried by pretrained features rather than by the kernel self-supervised machinery.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If KREPES's claims hold, self-supervised representation learning no longer requires a black-box deep network: a kernel model with a few thousand landmarks can be trained with the same losses at comparable accuracy.
  • Because the learned coefficients are the explanation, auditing a representation reduces to ranking landmarks by influence; the paper demonstrates this by exposing demographic proxies in Adult-1M without any supervision.
  • The framework is loss-agnostic, so new or custom SSL objectives can be plugged in and still inherit the interpretability and scaling benefits, as long as the loss is differentiable.
  • The reported parameter counts (0.5–0.9M versus 5–22M for the neural baselines) suggest kernel SSL could be a lighter-weight alternative for compute-limited settings, though the pretrained-kernel caveat matters.
  • The class-coverage metric κ, which tracks how few top landmarks span all classes, offers a training-free signal that correlates with downstream accuracy and could be used for model selection.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A fair test of the 'kernel SSL scales' claim would run KREPES with a randomly initialized eNTK and a frozen-backbone baseline; the paper only compares against a fully-trained pretrained network, so it does not isolate whether the pretrained features or the kernel machinery produce the gains.
  • The abstract promises a 'Feature Alignment Gap' metric that the main text never defines; the transparency claims would be easier to check if that metric were specified, since the body only defines influence scores and concept-alignment profiles.
  • The conceptual influence profile requires a labeled set of concept-positive and concept-negative examples (e.g., 500 'Sea' samples), which reintroduces the labeling cost that SSL is meant to avoid; extending this to zero-shot concept definitions would be a natural next step.
  • The observation that SSL objectives self-organize landmark norms by semantic class (small κ for high accuracy) hints at a general principle—that contrastive and joint-embedding losses allocate representational weight across categories even without labels—which, if confirmed, would make κ a useful diagnostic for representation quality.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes KREPES, a kernel-based representation-learning framework in which an SSL loss (SimCLR, BYOL, VICReg, Barlow Twins, etc.) is minimized over functions of the form f(x)=A^T k_x + gamma, where k_x are kernel evaluations against a small set of Nyström landmarks. A principal-component initialization, two preconditioning schemes, and randomized leverage-score landmark selection are introduced to make optimization practical. The paper claims that KREPES matches deep-network SSL accuracy on datasets up to 1M samples while using far fewer trainable parameters, and that the resulting representations are interpretable through representer landmarks and influence scores. The interpretability section also promises concept-level explanation via concept activation vectors. However, the empirical comparison is confounded for CIFAR-10, several advertised deliverables are absent from the main text, and the interpretability metrics are largely definitional rather than validated.

Significance. If the claims were fully supported, the paper would make a useful contribution: scaling kernel-based SSL objectives via Nyström approximation and using the resulting representer structure for post-hoc explanations is a natural and potentially valuable idea. The algorithmic components (eNTK + Nyström integration, principal-component initialization, GGN preconditioning for SSL losses) are concrete and could benefit practitioners. However, the current validation does not establish the central claims: the headline CIFAR-10 comparison is unfair, the scalability and bias-audit promises in the abstract are not implemented, and the influence-score interpretability is a restatement of the model's definition rather than an independently validated explanation. The significance is therefore conditional on a major revision that supplies fair baselines, complete experiments, and a reframed interpretability evaluation.

major comments (5)
  1. [§4, Table 1; App. B.2] The central empirical claim that KREPES performs comparably to deep SSL is not supported as stated. Appendix B.2 says that for CIFAR-10 the kernel is the eNTK of an ImageNet-pretrained ResNet34 and that the NN baseline is the same pretrained ResNet34. Table 1 then shows KREPES at 88.8–90.3% versus NN at 41.5–60.0%. If the baseline is actually trained from that initialization, the comparison conflates the pretrained feature geometry with the proposed SSL objective; if the baseline is trained from random initialization, KREPES receives an unacknowledged benefit of ImageNet features. In either case, the comparison does not isolate KREPES's contribution. The 'fewer parameters' statement is also misleading because the frozen eNTK ResNet34 backbone contributes roughly 22M parameters to the effective model. A fair comparison requires either a randomly initialized eNTK for CIFAR-10 or a frozen-f
  2. [Abstract and §5] The abstract promises deliverables that do not appear in the paper: (i) the Adult-1M demographic-proxy bias audit ('revealing an algorithmic bias ... where SSL uses demographic proxies for income'), (ii) scalability to ImageNet-1K, and (iii) the 'Feature Alignment Gap' metric. Section 5 and Appendix E contain only the CIFAR-100 sea example; no Adult-1M audit, no ImageNet-1K experiment, and no definition of 'Feature Alignment Gap' appear anywhere in the text. These are advertised contributions, not optional extras; they must either be implemented or removed from the claims.
  3. [§5.1, Eq. (3)] The proposed influence score is definitional rather than validated. Since f(x)=A^T k_x + gamma, the quantity iota_{l→t} = k(x_test, x_l) ||A_l|| is exactly the norm of the l-th representer term that composes the model output. Showing that high-scoring landmarks are close to the test input and have large learned coefficients is a restatement of the kernel model, not evidence that the explanation is meaningful to a user. The paper should compare against alternative attribution methods (e.g., leave-one-out, SHAP) or provide a quantitative user study or ground-truth-based evaluation of explanation quality. Relatedly, the class-coverage metric kappa in §5 uses ground-truth labels Y and y(x_l), so the 'label-free' characterization of the interpretability analysis is overstated.
  4. [§4, Table 1; App. F] The timing comparison is incomplete. Table 1 reports only the time per forward/backward pass multiplied by the number of epochs needed after the kernel is available. Appendix F states that computing the CIFAR-10 eNTK for ResNet34 takes less than 2 hours on four GPUs. KREPES is therefore not end-to-end 'seconds' training on CIFAR-10 when kernel construction is included. Because scalability is a central contribution, the total wall-clock including kernel construction should be reported, or the claims should be explicitly restricted to the optimization phase.
  5. [§4, Tables 1 and 2] All experimental tables report point estimates without error bars or number of seeds. The differences in Table 1 are often fractions of a percent (e.g., Adult-1M: 83.78 vs 83.72), and Table 2 compares preconditioners on similar small margins. Given the many hyperparameters tuned per dataset and loss (lambda, m, h, tau, VICReg coefficients, early stopping), the stability of these comparisons is unknown. At least three independent runs with standard deviations should be reported for the main claims.
minor comments (6)
  1. [Title and abstract] The manuscript's title in the provided text is 'Interpretable Kernel Representation Learning at Scale...', while the arXiv title is 'Interpretable Self-Supervised Learning via Representer Landmarks and Nyström Approximation'. The mismatch should be resolved in the camera-ready version.
  2. [§5.1, Eq. (3)] In the definition of Score(l→t, c), the term k(x_test, c_j) is likely a typo; it should presumably be k(x_test, x_l). The symbol c_j is not defined.
  3. [§5.1] The notation for global importance is inconsistent: earlier in Section 5 the landmark importance is defined as omega_l = ||a_l||_2 (row norm), but Section 5.1 refers to 'the landmark's corresponding column in the learned projection matrix ||A_l||'. Please clarify the exact indexing of A.
  4. [App. D, Table 5] The label 'Random Uniform' should be 'uniform random' or 'random sampling' to avoid implying a separate 'random' versus 'uniform' method.
  5. [§5, Figure 2] Figure 2 is referenced but not analyzed in the text beyond its caption. Please add a paragraph explaining what the figure demonstrates and how the influence scores should be read.
  6. [App. B.3] The text says 'maximum of 50 iterations' and later 'maximum number of epochs is between 60 and 80'; the units (iterations vs epochs) need to be reconciled.

Circularity Check

1 steps flagged

One self-definitional interpretability metric; the CIFAR-10 comparison shows a pretrained-feature confound but not a circular derivation.

specific steps
  1. self definitional [Section 5.1, Eq. defining ι_{l→t}; cf. Section 3 f(x)=A^T k_x+γ]
    "we define the influence of landmark x_l on a sample’s representation as the product of two factors: (i) local importance; how similar is the sample to the landmark k(x_test, x_l) (ii) global importance; how important is the landmark to the overall model captured by the norm of the landmark’s corresponding column in the learned projection matrix ||A_l|| as ι_{l→t} = k(x_test, x_l) · ||A_l||."

    In Section 3 the model is f(x)=A^T k_x+γ, so the contribution of landmark l to f(x_test) is exactly A_l k(x_test,x_l), whose norm is |k(x_test,x_l)|·||A_l|| = ι_{l→t}. Thus the proposed influence score is not an independent explanation of the representation; it is the magnitude of the landmark's own additive term in the fitted model output, by construction. The paper even states 'representer point interpretation is inherent to the nature of KREPES', confirming the metric is definitional rather than an empirical result that could be falsified.

full rationale

The core kernel derivation is self-contained: the representer-theorem reduction, Nyström approximation, and gradient-based optimization follow from standard external results, not from the paper's conclusions. The CIFAR-10 comparison in Table 1 is seriously confounded (Appendix B.2 uses the eNTK of an ImageNet-pretrained ResNet34 as kernel and the same pretrained ResNet34 as baseline, with an implausible 89–90% vs 41–60% gap), but this is an attribution/validity problem rather than a circular reduction: test accuracy is evaluated against an external benchmark and is not algebraically forced by the kernel definition. The abstract promises an Adult-1M demographic-proxy bias audit that does not appear in the full text, and Section 5's class-coverage metric κ uses ground-truth labels Y and y(x_l), contradicting the 'without supervision' claim; again, these are missing-support/overclaim issues, not circularity. Self-citations to Esser et al. supply loss definitions and an orthogonality constraint, but no unique theorem is imported, so they are not load-bearing in a circular way. The single genuine reduction is the influence score, which is the norm of the model's own term by construction. Hence a moderate score, not a higher one.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 4 invented entities

The central method is parameter-light (only A~ and gamma), which is honest. The main burden sits in hyperparameters tuned on labels, the unverified pretrained-kernel assumption, and the interpretability entities that are defined in terms of the model’s own outputs.

free parameters (6)
  • Regularization coefficient lambda (per loss, per dataset) = sampled log-uniform, e.g., [1e-5, 100]
    Tuned by random search on a labeled validation set for each loss-dataset combination (B.3); the paper acknowledges this compromises the unsupervised nature.
  • Number of landmarks m = 2000 for all datasets except MNIST
    Chosen by hand; no sensitivity analysis beyond landmark-selection ablation on MNIST (D).
  • Representation dimension h = 150 or 250 per loss
    From B.2: h depends on the loss function, no justification given.
  • SimCLR temperature tau = sampled from [1e-3, 100]
    A loss-specific hyperparameter tuned with labeled validation (B.3).
  • VICReg coefficients (tau, lambda, mu, nu) = sampled from [1e-5, 100]
    Four loss hyperparameters tuned per dataset (B.3).
  • Early stopping epochs/patience = max 60-80 epochs, patience 10-20
    Tuned per sweep (B.3).
axioms (4)
  • standard math Representer theorem applies to the SSL losses with Tikhonov regularization, restricting the optimum to the span of training points (Kimeldorf & Wahba 1970; Schölkopf et al. 2001).
    Invoked in Section 2 to justify the finite-dimensional parameterization; standard result, but its extension to the listed losses is asserted, not proven for every loss form.
  • domain assumption The eNTK of a pretrained network is a positive definite kernel adequate for SSL training.
    Used throughout experiments (Appendix A, B.2); the paper notes ‘we observed empirically that eNTKs by far have better performance’ but no correctness argument links the kernel to the SSL objective.
  • domain assumption Nyström approximation with m=2000 landmarks preserves sufficient information for SSL loss optimization.
    Motivated by Rudi et al. (2017) for KRR, but the paper does not extend the statistical guarantee to the nonconvex SSL objectives.
  • domain assumption The Adult-1M dataset is a valid expansion of Adult that preserves the demographic-proxy structure.
    The provided abstract’s bias claim depends on this dataset, but its construction is not described in the body.
invented entities (4)
  • Representer Landmarks no independent evidence
    purpose: Anchor points (Nyström landmarks) whose learned coefficients directly weight each sample’s representation, enabling interpretation.
    These are Nyström centers with a new name. The interpretative claim (that their influence decomposes a representation) is true by construction, so there is no independent falsifiable handle beyond the model’s own output.
  • Sample-Specific Influence Score iota_{l->t} no independent evidence
    purpose: Quantifies how much landmark l affects the representation of test point t as k(x_test, x_l) * ||A~_l||.
    It is a product of an already-defined kernel and the model’s own weights, so the explanation cannot be verified against anything outside the model.
  • Concept-Conditioned Influence Score Score(l->t,c) no independent evidence
    purpose: Combines CAV alignment with the influence score to attribute predictions to concepts.
    Depends on a linear SVC trained on representations, whose direction is itself model-derived. The qualitative gallery is the only support. Marked no independent evidence.
  • Feature Alignment Gap no independent evidence
    purpose: A metric named in the provided abstract to quantify transparency.
    It is listed in the provided abstract as a contribution and a metric name, but it never appears in the body; no definition is given.

pith-pipeline@v1.3.0-alltime-deepseek · 18966 in / 13225 out tokens · 147451 ms · 2026-08-04T14:32:52.975006+00:00 · methodology

0 comments
read the original abstract

Self-supervised learning (SSL) learns representations from massive unlabeled data, yet the resulting models typically operate as black boxes, necessitating domain-specific explanations. We introduce KREPES, a unified framework to analytically interpret the learned representations of SSL objectives, including SimCLR, BYOL, and VICReg. By bridging empirical neural tangent kernel approximations of neural networks with the Representer Theorem for kernels, we express the learned latent space directly via "Representer Landmarks", which are the representations of influential unlabeled training examples. We introduce novel metrics, "Sample-Specific Influence Score", "Concept-Conditioned Influence Score" and "Feature Alignment Gap", to quantify the transparency of the learned representations. KREPES enables direct audit of the latent space without supervision, for example, revealing an algorithmic bias in the Adult-1M dataset where SSL uses demographic proxies for income. Finally, to ensure scalability to benchmarks with 1M+ samples (ImageNet-1K, Adult-1M), KREPES introduces a novel Nystr\"om approximation-based analytical inference framework for SSL objectives.

Figures

Figures reproduced from arXiv: 2509.24467 by Debarghya Ghoshdastidar, Maedeh Zarvandi, Michael Timothy, Theresa Wasserer.

Figure 1
Figure 1. Figure 1: Eigenvalue spectrum comparison of A˜⊤A˜ under different preconditioners with BT. network predictions, and for a test point xt write output as Φ(xt, Θ∗ ) = Pn i=1 k(xt, xi , αi), where k(xt, xi , αi) = αif ⊤ i ft is the representer value of training point xi for xt, and ft denotes the pre￾activation of xt. The coefficients αi are derived from the training loss, making xi a representer point influencing the … view at source ↗
Figure 2
Figure 2. Figure 2: Sample-specific influential landmarks. First row, [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

15 extracted references · 5 linked inside Pith

  1. [9]

    Julien Mairal

    mlsys.org, 2019. Julien Mairal. End-to-end kernel learning with supervised convolutional kernel net- works. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett (eds.),Ad- vances in Neural Information Processing Systems, volume 29. Curran Associates, Inc.,

    2019

  2. [12]

    14 Under review as a conference paper at ICLR 2026 Ravid Shwartz-Ziv and Amitai Armon

    URLhttps://arxiv.org/abs/2003.02237. 14 Under review as a conference paper at ICLR 2026 Ravid Shwartz-Ziv and Amitai Armon. Tabular data: Deep learning is not all you need.Inf. Fusion, 81:84–90, 2022. James B Simon, Maksis Knutins, Liu Ziyin, Daniel Geisz, Abraham J Fetterman, and Joshua Al- brecht. On the stepwise nature of self-supervised learning. InIn...

    Pith/arXiv arXiv 2003

  3. [14]

    Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and St´ephane Deny

    URLhttps://proceedings.neurips.cc/paper_files/paper/2018/ file/8a7129b8f3edd95b7d969dfc2c8e9d9d-Paper.pdf. Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and St´ephane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In Marina Meila and Tong Zhang (eds.),Proceedings of the 38th International Conference on Machine Learning, ICML 2021...

    2018

  4. [256]

    Loss Function Hyper-parametersFor the simple contrastive loss, spectral contrastive loss, BT, and KPCA, the regularization coefficientλis sampled from the range[10 −5,100]

    For early stopping, the maximum number of epochs is between 60 and 80 with a patience of 10 to 20. Loss Function Hyper-parametersFor the simple contrastive loss, spectral contrastive loss, BT, and KPCA, the regularization coefficientλis sampled from the range[10 −5,100]. For SimCLR, the temperature parameterτis sampled from the interval[10 −3,100]. For VI...

    2015

  5. [1411]

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J ´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin

    JMLR.org, 2016. Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J ´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10- 17, 2021, pp. 9630–9640. IEEE, 2021. Ting Chen, Simon Kornblith...

    Pith/arXiv arXiv 2016

  6. [1989]

    Arthur Jacot, Franck Gabriel, and Clement Hongler

    URLhttps://api.semanticscholar.org/CorpusID:120969358. Arthur Jacot, Franck Gabriel, and Clement Hongler. Neural tangent kernel: Convergence and gen- eralization in neural networks. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa- Bianchi, and R. Garnett (eds.),Advances in Neural Information Processing Systems, volume 31. Curran Associates, I...

    Pith/arXiv arXiv 2018

  7. [2008]

    Elias Frantar, Eldar Kurtic, and Dan Alistarh

    URLhttp://jmlr.org/papers/v9/fan08a.html. Elias Frantar, Eldar Kurtic, and Dan Alistarh. M-fac: Efficient matrix-free approxi- mations of second-order information. In M. Ranzato, A. Beygelzimer, Y . Dauphin, P.S. Liang, and J. Wortman Vaughan (eds.),Advances in Neural Information Pro- cessing Systems, volume 34, pp. 14873–14886. Curran Associates, Inc., 2...

    2021

  8. [2011]

    URLhttps://doi.org/10.1137/090771806

    doi: 10.1137/090771806. URLhttps://doi.org/10.1137/090771806. 12 Under review as a conference paper at ICLR 2026 Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. InProceedings of the IEEE International Conference on Computer Vision (ICCV), pp. 1026–1034, 201...

    doi:10.1137/090771806 2026

  9. [2014]

    Neil Mallinar, James B

    URLhttps://proceedings.neurips.cc/paper_files/paper/2014/ file/768fa80ce4990fe601f5b2e094950511-Paper.pdf. Neil Mallinar, James B. Simon, Amirhesam Abedsoltan, Parthe Pandit, Misha Belkin, and Preetum Nakkiran. Benign, tempered, or catastrophic: Toward a refined taxonomy of overfitting. In Sanmi Koyejo, S. Mohamed, A. Agarwal, Danielle Belgrave, K. Cho, a...

    Pith/arXiv arXiv 2014

  10. [2016]

    Julien Mairal, Piotr Koniusz, Zaid Harchaoui, and Cordelia Schmid

    URLhttps://proceedings.neurips.cc/paper_files/paper/2016/ file/fc8001f834f6a5f0561080d134d53d29-Paper.pdf. Julien Mairal, Piotr Koniusz, Zaid Harchaoui, and Cordelia Schmid. Convolutional kernel net- works. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger (eds.), Advances in Neural Information Processing Systems, volume 27. Curran...

    2016

  11. [2018]

    Behrooz Ghorbani, Song Mei, Theodor Misiakiewicz, and Andrea Montanari

    URLhttps://proceedings.neurips.cc/paper_files/paper/2018/ file/48000647b315f6f00f913caa757a70b3-Paper.pdf. Behrooz Ghorbani, Song Mei, Theodor Misiakiewicz, and Andrea Montanari. When do neural networks outperform kernel methods? In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin (eds.),Advances in Neural Info...

    arXiv 2018

  12. [2019]

    David Arthur and Sergei Vassilvitskii

    URLhttps://proceedings.neurips.cc/paper_files/paper/2019/ file/dbc4d84bfcfe2284ba11beffb853a8c4-Paper.pdf. David Arthur and Sergei Vassilvitskii. k-means++: the advantages of careful seeding. InProceed- ings of the Eighteenth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’07, pp. 1027–1035, USA, 2007. Society for Industrial and Applied Mathematic...

    2019

  13. [2020]

    Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar R ¨atsch, Sylvain Gelly, Bernhard Sch¨olkopf, and Olivier Bachem

    URLhttps://proceedings.neurips.cc/paper_files/paper/2020/ file/ad086f59924fffe0773f8d0ca22ea712-Paper.pdf. Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar R ¨atsch, Sylvain Gelly, Bernhard Sch¨olkopf, and Olivier Bachem. Challenging common assumptions in the unsupervised learn- ing of disentangled representations, 2019. URLhttps://arxiv.org/abs/181...

    Pith/arXiv arXiv 2020

  14. [2024]

    Pascal Mattia Esser, Maximilian Fleissner, and Debarghya Ghoshdastidar

    URLhttps://proceedings.iclr.cc/paper_files/paper/2024/file/ 7319b7561ffe5e2f6419acd4a2f52d6b-Paper-Conference.pdf. Pascal Mattia Esser, Maximilian Fleissner, and Debarghya Ghoshdastidar. Non-parametric repre- sentation learning with kernels. In Michael J. Wooldridge, Jennifer G. Dy, and Sriraam Natarajan (eds.),Thirty-Eighth AAAI Conference on Artificial ...

    2024

  15. [2195]

    Alexander Wei, Wei Hu, and Jacob Steinhardt

    PMLR, 2020. Alexander Wei, Wei Hu, and Jacob Steinhardt. More than a toy: Random matrix models pre- dict how real-world neural representations generalize. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato (eds.),Proceedings of the 39th Inter- national Conference on Machine Learning, volume 162 ofProceedings of ...

    2020