pith. sign in

arxiv: 2606.04429 · v1 · pith:4UF5SO5Lnew · submitted 2026-06-03 · 📊 stat.ML · cs.LG

Flatness and Generalization: Learning Multi-Index Models with Homogeneous Neural Networks

Harsh Vardhan , Hossein Taheri , Arya Mazumdar This is my paper

Pith reviewed 2026-06-28 04:30 UTC · model grok-4.3

classification 📊 stat.ML cs.LG
keywords flatnessgeneralizationmulti-index modelshomogeneous neural networksinterpolatorssingle-index modelsHessian tracepopulation loss
0
0 comments X

The pith

For data generated by sums of single-index models with low noise, any flattest interpolator of homogeneous neural networks achieves small population loss.

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

The paper examines whether flatness of neural network solutions, measured by the trace of the Hessian, reliably indicates good generalization even though network symmetries can alter flatness without changing losses. It focuses on the flattest interpolators, defined as those achieving the orderwise minimum flatness among all zero-training-loss solutions. The authors first identify a class of non-generalizing interpolators whose flatness cannot be reduced to the minimum level even after applying symmetries. For data exactly generated as a sum of single-index models with low approximation error and label noise, they prove that every such flattest interpolator must have small population loss, establishing a direct flatness-generalization link for 2-layer homogeneous networks.

Core claim

For learning an unknown multi-index model with 2-layer non-convex homogeneous neural networks, there is a connection between flatness and generalization that persists despite symmetries. There exists a natural class of non-generalizing interpolators whose flatness cannot be made closer to the flattest possible even using symmetries. For data generated by a sum of single-index models, if the approximation error and label noise are low, any flattest interpolator achieves small population loss. This holds for a large class of activations and realistic data distributions.

What carries the argument

The flattest interpolators, defined as those with orderwise minimum trace of the Hessian of the empirical loss among all interpolators.

If this is right

  • Non-generalizing interpolators exist whose flatness cannot reach the minimum level even after symmetries are applied.
  • Any flattest interpolator achieves small population loss when data follows the sum-of-single-index model with low error and noise.
  • The flatness-generalization connection applies across a large class of activations and realistic distributions.
  • The result gives a direct link between minimum flatness and generalization for this class of models and data.

Where Pith is reading between the lines

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

  • Optimization trajectories that reliably reach minimum-flatness solutions could be favored for generalization in multi-index settings.
  • Similar arguments might apply to other homogeneous architectures if the single-index decomposition structure is preserved.
  • Flatness could serve as a practical selection criterion among multiple interpolators when the data-generating process is close to a sum of single-index models.

Load-bearing premise

The data must be generated exactly as a sum of single-index models with low approximation error and label noise.

What would settle it

Observe a flattest interpolator (minimum-order Hessian trace) on low-noise sum-of-single-index data that nonetheless has large population loss.

Figures

Figures reproduced from arXiv: 2606.04429 by Arya Mazumdar, Harsh Vardhan, Hossein Taheri.

Figure 1
Figure 1. Figure 1: Proof Sketch of Thms. 1 and 2. an ℓp norm, we can obtain a tight lower bound on it via the ℓ2 norm of the minimum ℓ2 norm interpolator in Lemma 5. The remainder of the proof utilizes two different bounds on }amin,ℓ2 } 2 for all interpolators and bad interpolators respectively. For bad interpolators that are not aligned with the true direction, Θ‹ , the labels y are almost independent of the activation matr… view at source ↗
Figure 2
Figure 2. Figure 2: Flatness and population loss for learning single-index link functions close to activations. In all [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Flatness and population loss for learning sum of single-index link functions, each close to [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Flatness and population loss for special data distributions. For learning linear link with ReLU [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
read the original abstract

A common heuristic used to explain the generalization of first-order gradient methods on non-convex neural networks is that "flat interpolators generalize well" (Hochreiter and Schmidhuber, 1994; Keskar et al., 2017), where flatness can be measured by the trace of the Hessian of the empirical loss. However, Dinh et al. 2017) showed that, using symmetry of the network that can change flatness while keeping the population and empirical losses unchanged, any interpolator can be made sharper or flatter. This result makes the earlier heuristic statement vacuous. In this paper, we show that for learning an unknown multi-index model with $2$-layer non-convex homogeneous neural networks, there is a connection between flatness and generalization, despite the existence of symmetries. This connection pertains to the "flattest" interpolators, i.e., the interpolators that have orderwise minimum flatness among all interpolators. First, we show that there exists a natural class of non-generalizing interpolators whose flatness cannot be made closer to the flattest possible, even using symmetries. Second, we show that for data generated by a sum of single-index models, if the approximation error and label noise are low, any flattest interpolator achieves small population loss, i.e., the flattest interpolators always generalize. This establishes a direct link between flatness and generalization which applies to a large class of activations and realistic data distributions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 2 minor

Summary. The paper claims that, for 2-layer homogeneous neural networks learning multi-index models, symmetries do not render the flatness heuristic vacuous when attention is restricted to orderwise-minimal flatness (flattest interpolators). Specifically, non-generalizing interpolators exist whose flatness cannot be reduced to the minimum even after symmetries, and, when data are generated exactly as a sum of single-index models with low approximation error and label noise, every flattest interpolator achieves small population loss. The result is stated to hold for a large class of activations.

Significance. If the derivations hold, the work supplies a non-vacuous, symmetry-aware link between flatness (trace of the Hessian) and generalization for this concrete function class and data model. The restriction to orderwise-minimal flatness is a coherent way to evade Dinh et al. reparameterizations while still obtaining a positive statement. The explicit conditioning on the sum-of-single-index data model is stated up front, so the result is not overclaimed for arbitrary targets.

minor comments (2)
  1. The abstract states the data-generation assumption clearly, but the introduction or §2 should contain an explicit statement of how the multi-index model is formalized (e.g., the precise form of the target function and the homogeneity degree of the network) so that the scope of the “large class of activations” is immediately visible.
  2. Notation for the flatness measure (trace of the Hessian of the empirical loss) and for “orderwise minimum flatness” should be introduced once in a dedicated subsection rather than scattered across the technical sections.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive summary, significance assessment, and recommendation of minor revision. No major comments were raised in the report.

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper's central result is a conditional theorem: when data is generated exactly by a sum of single-index models with low approximation error and label noise, any orderwise-minimal-flatness interpolator achieves small population loss. This premise is stated explicitly in the abstract and is not derived from the conclusion; the derivation therefore does not reduce to a self-definition, a fitted quantity renamed as a prediction, or a load-bearing self-citation chain. The treatment of Dinh et al. symmetries via the orderwise-minimal-flatness restriction is a methodological choice that sidesteps reparameterization without circularity. No equations or steps in the provided text exhibit the enumerated circular patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract only; no explicit free parameters, axioms, or invented entities are stated in the available text.

pith-pipeline@v0.9.1-grok · 5809 in / 1063 out tokens · 49365 ms · 2026-06-28T04:30:41.431511+00:00 · methodology

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

86 extracted references · 7 canonical work pages

  1. [1]

    Stable Minima Cannot Overfit in Univariate Re

    Dan Qiao and Kaiqi Zhang and Esha Singh and Daniel Soudry and Yu-Xiang Wang , booktitle=. Stable Minima Cannot Overfit in Univariate Re

  2. [2]

    Proceedings of the 40th International Conference on Machine Learning , pages =

    The Implicit Regularization of Dynamical Stability in Stochastic Gradient Descent , author =. Proceedings of the 40th International Conference on Machine Learning , pages =. 2023 , volume =

    2023

  3. [3]

    On Linear Stability of SGD and Input-Smoothness of Neural Networks , volume =

    Ma, Chao and Ying, Lexing , booktitle =. On Linear Stability of SGD and Input-Smoothness of Neural Networks , volume =

  4. [4]

    How SGD Selects the Global Minima in Over-parameterized Learning: A Dynamical Stability Perspective , volume =

    Wu, Lei and Ma, Chao and E, Weinan , booktitle =. How SGD Selects the Global Minima in Over-parameterized Learning: A Dynamical Stability Perspective , volume =

  5. [5]

    Mulayoff, Rotem and Michaeli, Tomer and Soudry, Daniel , year =. The. Advances in

  6. [6]

    Characterizing

    Chemnitz, Dennis and Engel, Maximilian , month = sep, year =. Characterizing. doi:10.48550/arXiv.2407.20209 , publisher =

    work page doi:10.48550/arxiv.2407.20209

  7. [7]

    Implicit

    Nacson, Mor Shpigel and Ravichandran, Kavya and Srebro, Nathan and Soudry, Daniel , month = jun, year =. Implicit. Proceedings of the 39th

  8. [8]

    Advances in Neural Information Processing Systems , author =

    The alignment property of. Advances in Neural Information Processing Systems , author =. 2022 , pages =

    2022

  9. [9]

    Mulayoff, Rotem and Michaeli, Tomer , month = jun, year =. Exact. Proceedings of

  10. [10]

    2023 , cdate=

    Mor Shpigel Nacson and Rotem Mulayoff and Greg Ongie and Tomer Michaeli and Daniel Soudry , title=. 2023 , cdate=

    2023

  11. [11]

    Journal of the American Statistical Association , volume =

    Peter L Bartlett and Michael I Jordan and Jon D McAuliffe , title =. Journal of the American Statistical Association , volume =. 2006 , publisher =

    2006

  12. [12]

    Proceedings of Thirty Fifth Conference on Learning Theory , pages =

    Stochastic linear optimization never overfits with quadratically-bounded losses on general data , author =. Proceedings of Thirty Fifth Conference on Learning Theory , pages =. 2022 , volume =

    2022

  13. [13]

    Self-concordant analysis for logistic regression

    Bach, Francis. Self-concordant analysis for logistic regression. Electron. J. Stat

  14. [14]

    2025 , eprint=

    Flat Minima and Generalization: Insights from Stochastic Convex Optimization , author=. 2025 , eprint=

    2025

  15. [15]

    9th International Conference on Learning Representations,

    Pierre Foret and Ariel Kleiner and Hossein Mobahi and Behnam Neyshabur , title =. 9th International Conference on Learning Representations,

  16. [16]

    International Conference on Learning Representations , year=

    Fantastic Generalization Measures and Where to Find Them , author=. International Conference on Learning Representations , year=

  17. [17]

    Proceedings of the 40th International Conference on Machine Learning , pages =

    A Modern Look at the Relationship between Sharpness and Generalization , author =. Proceedings of the 40th International Conference on Machine Learning , pages =. 2023 , volume =

    2023

  18. [18]

    The Eleventh International Conference on Learning Representations , year=

    How Sharpness-Aware Minimization Minimizes Sharpness? , author=. The Eleventh International Conference on Learning Representations , year=

  19. [19]

    Information and Inference: A Journal of the IMA , volume =

    Ding, Lijun and Drusvyatskiy, Dmitriy and Fazel, Maryam and Harchaoui, Zaid , title =. Information and Inference: A Journal of the IMA , volume =. 2024 , month =

    2024

  20. [20]

    Stable Minima of Re

    Tongtong Liang and Dan Qiao and Yu-Xiang Wang and Rahul Parhi , booktitle=. Stable Minima of Re

  21. [21]

    , keywords =

    Lee, John M. , keywords =

  22. [22]

    Proceedings of the National Academy of Sciences , volume =

    Mikhail Belkin and Daniel Hsu and Siyuan Ma and Soumik Mandal , title =. Proceedings of the National Academy of Sciences , volume =. 2019 , doi =

    2019

  23. [23]

    Wainwright, Martin J. , year=. High-Dimensional Statistics: A Non-Asymptotic Viewpoint , publisher=

  24. [24]

    and Mendelson, Shahar , title =

    Bartlett, Peter L. and Mendelson, Shahar , title =. J. Mach. Learn. Res. , month = mar, pages =. 2003 , publisher =

    2003

  25. [25]

    Statistical Science , number =

    Joan Bruna and Daniel Hsu , title =. Statistical Science , number =

  26. [26]

    , year =

    Vapnik, Vladimir N. , year =. The Nature of Statistical Learning Theory , ISBN =. doi:10.1007/978-1-4757-3264-1 , publisher =

    work page doi:10.1007/978-1-4757-3264-1

  27. [27]

    2021 , issue_date =

    Zhang, Chiyuan and Bengio, Samy and Hardt, Moritz and Recht, Benjamin and Vinyals, Oriol , title =. 2021 , issue_date =. doi:10.1145/3446776 , journal =

    work page doi:10.1145/3446776 2021

  28. [28]

    Zico , booktitle =

    Nagarajan, Vaishnavh and Kolter, J. Zico , booktitle =. Uniform convergence may be unable to explain generalization in deep learning , year =

  29. [29]

    International Conference on Learning Representations , year=

    The Implicit Bias of Gradient Descent on Separable Data , author=. International Conference on Learning Representations , year=

  30. [30]

    Proceedings of the 34th International Conference on Machine Learning , pages =

    Sharp Minima Can Generalize For Deep Nets , author =. Proceedings of the 34th International Conference on Machine Learning , pages =. 2017 , volume =

    2017

  31. [31]

    Proceedings of Thirty Third Conference on Learning Theory , pages =

    Learning Polynomials in Few Relevant Dimensions , author =. Proceedings of Thirty Third Conference on Learning Theory , pages =. 2020 , volume =

    2020

  32. [32]

    Lee and Taiji Suzuki , booktitle=

    Kazusato Oko and Denny Wu and Jason D. Lee and Taiji Suzuki , booktitle=. Neural network learns low-dimensional polynomials with

  33. [33]

    Sub‐Weibull distributions: Generalizing sub‐Gaussian and sub‐Exponential properties to heavier tailed distributions , volume=

    Vladimirova, Mariia and Girard, Stéphane and Nguyen, Hien and Arbel, Julyan , year=. Sub‐Weibull distributions: Generalizing sub‐Gaussian and sub‐Exponential properties to heavier tailed distributions , volume=. Stat , publisher=. doi:10.1002/sta4.318 , number=

    work page doi:10.1002/sta4.318

  34. [34]

    Electronic Journal of Probability , number =

    Friedrich G. Electronic Journal of Probability , number =. 2021 , doi =

    2021

  35. [35]

    2024 , eprint=

    Uniform Hanson-Wright Type Deviation Inequalities for -Subexponential Random Vectors , author=. 2024 , eprint=

    2024

  36. [36]

    Information and Inference: A Journal of the IMA , volume =

    Kuchibhotla, Arun Kumar and Chakrabortty, Abhishek , title =. Information and Inference: A Journal of the IMA , volume =. 2022 , month =

    2022

  37. [37]

    Some Notes on Concentration for -Subexponential Random Variables

    Sambale, Holger. Some Notes on Concentration for -Subexponential Random Variables. High Dimensional Probability IX. 2023

    2023

  38. [38]

    and Shakarchi, Rami , title =

    Stein, Elias M. and Shakarchi, Rami , title =. 2011 , address =

    2011

  39. [39]

    Mathematics , VOLUME =

    Zhang, Huiming and Wei, Haoyu , TITLE =. Mathematics , VOLUME =. 2022 , NUMBER =

    2022

  40. [40]

    Известия Российской академии наук

    Uber die abgrenzung der eigenwerte einer matrix , author=. Известия Российской академии наук. Серия математическая , number=. 1931 , publisher=

    1931

  41. [41]

    Electronic Communications in Probability , number =

    Mark Rudelson and Roman Vershynin , title =. Electronic Communications in Probability , number =. 2013 , doi =

    2013

  42. [42]

    2024 , eprint=

    Expressivity and Approximation Properties of Deep Neural Networks with ReLU ^k Activation , author=. 2024 , eprint=

    2024

  43. [43]

    Path-SGD: Path-Normalized Optimization in Deep Neural Networks , volume =

    Neyshabur, Behnam and Salakhutdinov, Russ R and Srebro, Nati , booktitle =. Path-SGD: Path-Normalized Optimization in Deep Neural Networks , volume =

  44. [44]

    Transactions on Machine Learning Research , issn=

    Symmetry in Neural Network Parameter Spaces , author=. Transactions on Machine Learning Research , issn=

  45. [45]

    2022 , issn =

    High-order approximation rates for shallow neural networks with cosine and ReLUk activation functions , journal =. 2022 , issn =. doi:https://doi.org/10.1016/j.acha.2021.12.005 , author =

    work page doi:10.1016/j.acha.2021.12.005 2022

  46. [46]

    Optimal Rates of Approximation by Shallow ReLU ^k Neural Networks and Applications to Nonparametric Regression , volume =

    Yang, Yunfei and Zhou, Ding-Xuan , year =. Optimal Rates of Approximation by Shallow ReLU ^k Neural Networks and Applications to Nonparametric Regression , volume =. Constructive Approximation , publisher =. doi:10.1007/s00365-024-09679-z , number =

    work page doi:10.1007/s00365-024-09679-z

  47. [47]

    Proceedings of Thirty Fifth Conference on Learning Theory , pages =

    Neural Networks can Learn Representations with Gradient Descent , author =. Proceedings of Thirty Fifth Conference on Learning Theory , pages =. 2022 , volume =

    2022

  48. [48]

    Lee , booktitle=

    Alex Damian and Eshaan Nichani and Rong Ge and Jason D. Lee , booktitle=. Smoothing the Landscape Boosts the Signal for

  49. [49]

    International Conference on Learning Representations , year=

    Gradient Descent on Neural Networks Typically Occurs at the Edge of Stability , author=. International Conference on Learning Representations , year=

  50. [50]

    The Fourteenth International Conference on Learning Representations , year=

    Gradient Descent with Large Step Sizes: Chaos and Fractal Convergence Region , author=. The Fourteenth International Conference on Learning Representations , year=

  51. [51]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Implicit Bias of Gradient Descent for Logistic Regression at the Edge of Stability , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

  52. [52]

    Proceedings of the 37th International Conference on Machine Learning , pages =

    Unique Properties of Flat Minima in Deep Networks , author =. Proceedings of the 37th International Conference on Machine Learning , pages =. 2020 , volume =

    2020

  53. [53]

    International Conference on Learning Representations , year=

    On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima , author=. International Conference on Learning Representations , year=

  54. [54]

    Simplifying neural nets by discovering flat minima , year =

    Hochreiter, Sepp and Schmidhuber, J\". Simplifying neural nets by discovering flat minima , year =. Proceedings of the 8th International Conference on Neural Information Processing Systems , pages =

  55. [55]

    Proceedings of the 38th International Conference on Machine Learning , pages =

    ASAM: Adaptive Sharpness-Aware Minimization for Scale-Invariant Learning of Deep Neural Networks , author =. Proceedings of the 38th International Conference on Machine Learning , pages =. 2021 , volume =

    2021

  56. [56]

    Relative Flatness and Generalization , volume =

    Petzka, Henning and Kamp, Michael and Adilova, Linara and Sminchisescu, Cristian and Boley, Mario , booktitle =. Relative Flatness and Generalization , volume =

  57. [57]

    Entropy-

    Dziugaite, Gintare Karolina and Roy, Daniel , booktitle =. Entropy-. 2018 , editor =

    2018

  58. [58]

    doi:10.1088/1742-5468/ab39d9 , year =

    Chaudhari, Pratik and Choromanska, Anna and Soatto, Stefano and LeCun, Yann and Baldassi, Carlo and Borgs, Christian and Chayes, Jennifer and Sagun, Levent and Zecchina, Riccardo , title =. doi:10.1088/1742-5468/ab39d9 , year =

    work page doi:10.1088/1742-5468/ab39d9

  59. [59]

    2025 , eprint=

    Generalization Below the Edge of Stability: The Role of Data Geometry , author=. 2025 , eprint=

    2025

  60. [60]

    2025 , eprint=

    Does Flatness imply Generalization for Logistic Loss in Univariate Two-Layer ReLU Network? , author=. 2025 , eprint=

    2025

  61. [61]

    The Eleventh International Conference on Learning Representations , year=

    Loss Landscapes are All You Need: Neural Network Generalization Can Be Explained Without the Implicit Bias of Gradient Descent , author=. The Eleventh International Conference on Learning Representations , year=

  62. [62]

    2023 , eprint=

    FAM: Relative Flatness Aware Minimization , author=. 2023 , eprint=

    2023

  63. [63]

    Proceedings of The 36th International Conference on Algorithmic Learning Theory , pages =

    A PAC-Bayesian Link Between Generalisation and Flat Minima , author =. Proceedings of The 36th International Conference on Algorithmic Learning Theory , pages =. 2025 , volume =

    2025

  64. [64]

    Journal of Machine Learning Research , year =

    Gerard Ben Arous and Reza Gheissari and Aukosh Jagannath , title =. Journal of Machine Learning Research , year =

  65. [65]

    2004 , publisher=

    Convex optimization , author=. 2004 , publisher=

    2004

  66. [66]

    Normalized Flat Minima: Exploring Scale Invariant Definition of Flat Minima for Neural Networks Using

    Tsuzuku, Yusuke and Sato, Issei and Sugiyama, Masashi , booktitle =. Normalized Flat Minima: Exploring Scale Invariant Definition of Flat Minima for Neural Networks Using. 2020 , volume =

    2020

  67. [67]

    Concentration Inequalities for Statistical Inference , volume=

    Zhang, Huiming and Chen, Songxi , year=. Concentration Inequalities for Statistical Inference , volume=. Communications in Mathematical Research , publisher=

  68. [68]

    Proceedings of Thirty Seventh Conference on Learning Theory , pages =

    Learning sum of diverse features: computational hardness and efficient gradient-based training for ridge combinations , author =. Proceedings of Thirty Seventh Conference on Learning Theory , pages =. 2024 , volume =

    2024

  69. [69]

    Bell System Technical Journal , volume =

    Slepian, David , title =. Bell System Technical Journal , volume =

  70. [70]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    What is the Inductive Bias of Flatness Regularization? A Study of Deep Matrix Factorization Models , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

  71. [71]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Provable Guarantees for Nonlinear Feature Learning in Three-Layer Neural Networks , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

  72. [72]

    Journal of Machine Learning Research , year =

    Steven Diamond and Stephen Boyd , title =. Journal of Machine Learning Research , year =

  73. [73]

    Journal of Control and Decision , year =

    Agrawal, Akshay and Verschueren, Robin and Diamond, Steven and Boyd, Stephen , title =. Journal of Control and Decision , year =

  74. [74]

    The MOSEK optimization toolbox for Python, version 11.1 , year =

  75. [75]

    Bartlett and Philip M

    Peter L. Bartlett and Philip M. Long and Gábor Lugosi and Alexander Tsigler , title =. Proceedings of the National Academy of Sciences , volume =

  76. [76]

    Laurent and P

    B. Laurent and P. Massart , title =. The Annals of Statistics , number =

  77. [77]

    Proceedings of Thirty Fifth Conference on Learning Theory , pages =

    Benign Overfitting without Linearity: Neural Network Classifiers Trained by Gradient Descent for Noisy Linear Data , author =. Proceedings of Thirty Fifth Conference on Learning Theory , pages =. 2022 , volume =

    2022

  78. [78]

    Journal of the American Statistical Association , volume =

    Ker-Chau Li , title =. Journal of the American Statistical Association , volume =. 1991 , publisher =

    1991

  79. [79]

    Proceedings of Thirty Sixth Conference on Learning Theory , pages =

    SGD learning on neural networks: leap complexity and saddle-to-saddle dynamics , author =. Proceedings of Thirty Sixth Conference on Learning Theory , pages =. 2023 , volume =

    2023

  80. [80]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    The Generative Leap: Tight Sample Complexity for Efficiently Learning Gaussian Multi-Index Models , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

Showing first 80 references.