pith. sign in

arxiv: 2606.10089 · v1 · pith:2PR7UVZInew · submitted 2026-06-08 · 💻 cs.LG · cs.AI

A Theory on Flow Matching with Neural Networks

Yihan He , Qishuo Yin , Yuan Cao , Jianqing Fan , Han Liu This is my paper

Pith reviewed 2026-06-27 16:57 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords flow matchingneural networksconvergence guaranteesgeneralization boundsWasserstein distancevelocity fieldsoverparameterized ReLU networksgenerative modeling
0
0 comments X

The pith

Overparameterized two-layer ReLU networks converge under gradient descent when trained to match conditional velocity fields, yielding Wasserstein guarantees on generated samples.

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

The paper develops theoretical guarantees for flow matching models where a neural network parameterizes the conditional velocity field that transports samples from noise to data. It shows gradient descent converges in the overparameterized regime for two-layer ReLU networks and derives generalization bounds on the matching objective. These bounds imply that the induced flow produces samples whose distribution is close to the target in Wasserstein distance. The analysis adapts a multi-task representation learning bound for unbounded losses, which supports the velocity-field results. Experiments on synthetic data and image benchmarks confirm the predicted behavior.

Core claim

We establish convergence guarantees for gradient descent in the over-parameterized 2-layered ReLU neural network regime. We derive generalization bounds for the conditional velocity-field matching objective. Building on these results, we provide Wasserstein-distance guarantees for the samples generated by the induced flow. Our analysis is based on a generalization bound for multi-task representation learning with unbounded losses.

What carries the argument

The conditional velocity-field matching objective, which trains the network to approximate the time-dependent velocity that maps noise distributions to data distributions under the flow.

If this is right

  • Gradient descent reaches a solution with controlled error for the velocity field approximation.
  • The generated flow produces samples with provably bounded Wasserstein distance to the data distribution.
  • The same multi-task bound yields generalization results for the flow matching loss.
  • The guarantees extend to both synthetic distributions and real image datasets under the stated conditions.

Where Pith is reading between the lines

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

  • Similar analysis could apply to other continuous normalizing flow variants if their objectives admit comparable multi-task reductions.
  • The Wasserstein guarantees suggest that early stopping or regularization choices can be guided by the derived rates rather than cross-validation alone.
  • Extensions to deeper or wider networks would require checking whether the overparameterization assumptions scale without introducing new error terms.

Load-bearing premise

The multi-task representation learning bound for unbounded losses applies directly to the conditional velocity-field matching objective.

What would settle it

Training runs where the empirical Wasserstein distance between generated and target samples remains larger than the paper's derived bound after convergence in the stated overparameterized regime.

Figures

Figures reproduced from arXiv: 2606.10089 by Han Liu, Jianqing Fan, Qishuo Yin, Yihan He, Yuan Cao.

Figure 1
Figure 1. Figure 1: Global view of the 5 · 3 · 10 = 150 cell sweep at ntrain = 500, ntest = 5000. Each curve is one (m, η) pair; color encodes η (10−4 darkest, 10−2 lightest) and marker/linestyle encodes width m. Values at the final iterate t = T = 500; solid lines show the mean and shaded bands show ±1 standard deviation over 10 independent random seeds. (a) Terminal training loss grows with d and decreases with η (Theorem 3… view at source ↗
Figure 2
Figure 2. Figure 2: 8 × 8 pixel images generated in PCA code space (d = 7) and inverse-transformed; terminal iterate t = T = 500, cell (m, η) = (512, 10−2 ), ntrain = 500. Left: MNIST samples are recognizable as handwritten digits. Right: Fashion-MNIST samples display distinct garment categories with the smoothness characteristic of a 7-component PCA basis. Full per-cell reconstruction grids are in Appendix G.4.2. and sweep m… view at source ↗
Figure 3
Figure 3. Figure 3: Training loss L(θ (t) ) along the gradient-descent trajectory for every (m, η) in the grid, ntrain = 500, ntest = 5000. Within each panel, color encodes the ambient dimension d ∈ {5, 10, . . . , 50} (color bar); curves show the mean over 10 independent random seeds. All 15 panels exhibit geometric decay in t on the log-scale, in line with Theorem 3.2. 58 [PITH_FULL_IMAGE:figures/full_fig_p058_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sliced Wasserstein-1 distance W1(P1, Pxb (t) 1 ) along the gradient-descent trajectory for every (m, η) in the grid, ntrain = 500, ntest = 5000. Within each panel, color encodes the ambient dimension d ∈ {5, 10, . . . , 50} (color bar); curves show the mean over 10 independent random seeds. The distance drops monotonically along the trajectory and saturates at a value increasing in d, in line with Theorem … view at source ↗
Figure 5
Figure 5. Figure 5: Training dynamics on MNIST (left) and Fashion-MNIST (right), [PITH_FULL_IMAGE:figures/full_fig_p061_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: 8 × 8 grids of images generated by Algorithm 2 at the terminal iterate t = T = 500 for every (m, η) cell on MNIST, ntrain = 500. Rows correspond to widths m ∈ {128, 256, 512, 1024}; columns to step sizes η ∈ {10−4 , 10−3 , 10−2}. Sample quality improves with η; the effect of m at fixed η is small. 62 [PITH_FULL_IMAGE:figures/full_fig_p062_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: 8×8 grids of images generated by Algorithm 2 at the terminal iterate t = T = 500 for every (m, η) cell on Fashion-MNIST, ntrain = 500. Layout identical to [PITH_FULL_IMAGE:figures/full_fig_p063_7.png] view at source ↗
read the original abstract

In this work, we develop theoretical foundation for flow matching with neural-network-parameterized conditional velocity fields. We establish convergence guarantees for gradient descent in the over-parameterized 2-layered ReLU neural network regime. We derive generalization bounds for the conditional velocity-field matching objective. Building on these results, we provide Wasserstein-distance guarantees for the samples generated by the induced flow. Our analysis is based on generalization bound for multi-task representation learning with unbounded losses, which may be of independent interest beyond flow-based generative modeling. These theoretical results are validated through extensive experiments on both synthetic and real-world image benchmarks.

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

2 major / 2 minor

Summary. The paper develops theoretical foundations for flow matching with neural-network-parameterized conditional velocity fields. It claims convergence guarantees for gradient descent in the over-parameterized 2-layer ReLU regime, generalization bounds for the conditional velocity-field matching objective derived from a multi-task representation learning bound with unbounded losses, and resulting Wasserstein-distance guarantees for samples from the induced flow. Results are supported by experiments on synthetic and real-world image benchmarks.

Significance. If the mapping of the conditional velocity-field objective to the multi-task bound is valid and the derivations hold, the work would supply useful convergence and generalization theory for flow matching, a prominent class of generative models. The multi-task bound with unbounded losses is presented as potentially of independent interest.

major comments (2)
  1. [Abstract and theoretical analysis] The central claims (GD convergence, generalization of the matching objective, and Wasserstein guarantees) rest on applying the multi-task representation learning bound with unbounded losses to the conditional velocity-field matching objective, yet the manuscript supplies no explicit verification that the objective satisfies the bound's assumptions on loss structure, unboundedness handling, task decomposition, or representation-learning setup (see abstract and the theoretical analysis sections).
  2. [Theoretical analysis] The conditional nature of the velocity field (conditioned on data or time) must map onto the multi-task framework for the bound to apply directly; without a concrete check of this mapping or the over-parameterized ReLU regime, the applicability of the bound remains unestablished and undermines the downstream Wasserstein guarantees.
minor comments (2)
  1. [Abstract] The abstract asserts the existence of proofs and bounds but provides no derivation steps, assumption lists, or error-bar details, making it difficult to assess the mathematics even at a high level.
  2. [Experiments] Experiment section should include specific metrics, baselines, and quantitative results to support the validation claims on image benchmarks.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful review and constructive feedback on our manuscript. We address the major comments point by point below and will revise the manuscript to improve clarity on the mapping to the multi-task bound.

read point-by-point responses

  1. Referee: [Abstract and theoretical analysis] The central claims (GD convergence, generalization of the matching objective, and Wasserstein guarantees) rest on applying the multi-task representation learning bound with unbounded losses to the conditional velocity-field matching objective, yet the manuscript supplies no explicit verification that the objective satisfies the bound's assumptions on loss structure, unboundedness handling, task decomposition, or representation-learning setup (see abstract and the theoretical analysis sections).

    Authors: We agree that an explicit, systematic verification of the assumptions would strengthen the presentation and make the applicability clearer. While the theoretical analysis sections frame the conditional velocity-field objective within the multi-task representation learning setup (with tasks corresponding to conditioning on time and data), we acknowledge that a dedicated check listing each assumption (loss structure, unboundedness handling via truncation or moment conditions, task decomposition, and the 2-layer ReLU overparameterized regime) is not presented as a single consolidated verification. In the revision we will add a new subsection that performs this explicit mapping and assumption check. revision: yes

  2. Referee: [Theoretical analysis] The conditional nature of the velocity field (conditioned on data or time) must map onto the multi-task framework for the bound to apply directly; without a concrete check of this mapping or the over-parameterized ReLU regime, the applicability of the bound remains unestablished and undermines the downstream Wasserstein guarantees.

    Authors: We concur that the conditional structure requires an explicit mapping to justify direct application of the bound. In the revised manuscript we will add a concrete construction: we define tasks as pairs (t, x_0) where t is discretized time and x_0 indexes data samples, with the shared representation learned by the 2-layer ReLU network satisfying the overparameterization conditions of the bound. This explicit mapping will be inserted prior to the generalization and Wasserstein results to ensure the chain of implications is fully justified. revision: yes

Circularity Check

0 steps flagged

No circularity: analysis applies external multi-task generalization bound to flow-matching objective

full rationale

The paper states that its convergence guarantees, generalization bounds for the conditional velocity-field matching objective, and Wasserstein guarantees build on a generalization bound for multi-task representation learning with unbounded losses. This bound is described as potentially of independent interest beyond the present work. No quoted equation or derivation in the provided abstract reduces any claimed result to a fitted parameter, self-definition, or self-citation chain internal to the flow-matching analysis. The central claims therefore rest on an externally applicable bound rather than on quantities defined or fitted inside the paper itself.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claims depend on the applicability of an external generalization bound for multi-task representation learning with unbounded losses; no free parameters or new invented entities are introduced in the abstract.

axioms (1)
  • domain assumption A generalization bound for multi-task representation learning with unbounded losses holds and transfers to the flow-matching velocity-field objective
    The abstract states that the analysis is based on this bound.

pith-pipeline@v0.9.1-grok · 5624 in / 1192 out tokens · 21837 ms · 2026-06-27T16:57:34.790745+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

300 extracted references · 44 linked inside Pith

  1. [1]

    International Conference on Machine Learning , pages=

    A mean field analysis of deep resnet and beyond: Towards provably optimization via overparameterization from depth , author=. International Conference on Machine Learning , pages=. 2020 , organization=

    2020

  2. [2]

    2021 , eprint=

    On the Global Convergence of Gradient Descent for multi-layer ResNets in the mean-field regime , author=. 2021 , eprint=

    2021

  3. [3]

    Journal of Machine Learning Research , volume=

    Overparameterization of deep resnet: Zero loss and mean-field analysis , author=. Journal of Machine Learning Research , volume=

  4. [4]

    Advances in Neural Information Processing Systems , volume=

    Global convergence in training large-scale transformers , author=. Advances in Neural Information Processing Systems , volume=

  5. [5]

    Conference on learning theory , pages=

    Modeling from features: a mean-field framework for over-parameterized deep neural networks , author=. Conference on learning theory , pages=. 2021 , organization=

    2021

  6. [6]

    their Induced Kernel , author=

    Regularization Matters: Generalization and Optimization of Neural Nets v.s. their Induced Kernel , author=. Advances in Neural Information Processing Systems , year=

  7. [7]

    Proceedings of the National Academy of Sciences , volume=

    A mean field view of the landscape of two-layer neural networks , author=. Proceedings of the National Academy of Sciences , volume=. 2018 , publisher=

    2018

  8. [8]

    Advances in neural information processing systems , pages=

    On the global convergence of gradient descent for over-parameterized models using optimal transport , author=. Advances in neural information processing systems , pages=

  9. [9]

    Advances in Neural Information Processing Systems , year=

    On exact computation with an infinitely wide neural net , author=. Advances in Neural Information Processing Systems , year=

  10. [10]

    Advances in Neural Information Processing Systems , year=

    On Lazy Training in Differentiable Programming , author=. Advances in Neural Information Processing Systems , year=

  11. [11]

    Advances in Neural Information Processing Systems , year=

    Wide neural networks of any depth evolve as linear models under gradient descent , author=. Advances in Neural Information Processing Systems , year=

  12. [12]

    International Conference on Learning Representations , year=

    Gradient Descent Provably Optimizes Over-parameterized Neural Networks , author=. International Conference on Learning Representations , year=

  13. [13]

    Advances in Neural Information Processing Systems , pages=

    Learning overparameterized neural networks via stochastic gradient descent on structured data , author=. Advances in Neural Information Processing Systems , pages=

  14. [14]

    International Conference on Machine Learning , pages=

    A Convergence Theory for Deep Learning via Over-Parameterization , author=. International Conference on Machine Learning , pages=

  15. [15]

    Gradient descent optimizes over-parameterized deep ReLU networks

    Zou, Difan and Cao, Yuan and Zhou, Dongruo and Gu, Quanquan. Gradient descent optimizes over-parameterized deep ReLU networks. Machine Learning. 2019

    2019

  16. [16]

    Advances in Neural Information Processing Systems , year=

    Learning and Generalization in Overparameterized Neural Networks, Going Beyond Two Layers , author=. Advances in Neural Information Processing Systems , year=

  17. [17]

    Langley , title =

    P. Langley , title =. Proceedings of the 17th International Conference on Machine Learning (ICML 2000) , address =. 2000 , pages =

    2000

  18. [18]

    T. M. Mitchell. The Need for Biases in Learning Generalizations. 1980

    1980

  19. [19]

    M. J. Kearns , title =

  20. [20]

    Machine Learning: An Artificial Intelligence Approach, Vol. I. 1983

    1983

  21. [21]

    R. O. Duda and P. E. Hart and D. G. Stork. Pattern Classification. 2000

    2000

  22. [22]

    Suppressed for Anonymity , author=

  23. [23]

    Newell and P

    A. Newell and P. S. Rosenbloom. Mechanisms of Skill Acquisition and the Law of Practice. Cognitive Skills and Their Acquisition. 1981

    1981

  24. [24]

    A. L. Samuel. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development. 1959

    1959

  25. [25]

    Advances in neural information processing systems , volume=

    Understanding double descent requires a fine-grained bias-variance decomposition , author=. Advances in neural information processing systems , volume=

  26. [26]

    Advances in neural information processing systems , volume=

    Nonlinear random matrix theory for deep learning , author=. Advances in neural information processing systems , volume=

  27. [27]

    Random Matrices: Theory and Applications , volume=

    The spectrum of random inner-product kernel matrices , author=. Random Matrices: Theory and Applications , volume=. 2013 , publisher=

    2013

  28. [28]

    2012 , publisher=

    Topics in random matrix theory , author=. 2012 , publisher=

    2012

  29. [29]

    The Annals of Statistics , volume=

    Surprises in high-dimensional ridgeless least squares interpolation , author=. The Annals of Statistics , volume=. 2022 , publisher=

    2022

  30. [30]

    Journal of Statistical Mechanics: Theory and Experiment , volume=

    Scaling description of generalization with number of parameters in deep learning , author=. Journal of Statistical Mechanics: Theory and Experiment , volume=. 2020 , publisher=

    2020

  31. [31]

    arXiv preprint arXiv:1912.07242 , year=

    More data can hurt for linear regression: Sample-wise double descent , author=. arXiv preprint arXiv:1912.07242 , year=

    arXiv 1912

  32. [32]

    International Conference on Machine Learning , pages=

    Double trouble in double descent: Bias and variance (s) in the lazy regime , author=. International Conference on Machine Learning , pages=. 2020 , organization=

    2020

  33. [33]

    Physical Review Letters , volume=

    Eigenvalues of covariance matrices: Application to neural-network learning , author=. Physical Review Letters , volume=. 1991 , publisher=

    1991

  34. [34]

    Journal of Physics A: Mathematical and General , volume=

    Generalization in a linear perceptron in the presence of noise , author=. Journal of Physics A: Mathematical and General , volume=. 1992 , publisher=

    1992

  35. [35]

    2001 , publisher=

    Statistical mechanics of learning , author=. 2001 , publisher=

    2001

  36. [36]

    arXiv preprint arXiv:2003.01897 , year=

    Optimal regularization can mitigate double descent , author=. arXiv preprint arXiv:2003.01897 , year=

    arXiv 2003

  37. [37]

    Advances in Neural Information Processing Systems , volume=

    Triple descent and the two kinds of overfitting: Where & why do they appear? , author=. Advances in Neural Information Processing Systems , volume=

  38. [38]

    Advances in Neural Information Processing Systems , volume=

    Multiple descent: Design your own generalization curve , author=. Advances in Neural Information Processing Systems , volume=

  39. [39]

    2011 , publisher=

    Random fields on the sphere: representation, limit theorems and cosmological applications , author=. 2011 , publisher=

    2011

  40. [40]

    , author=

    Spherical-homoscedastic distributions: The equivalency of spherical and normal distributions in classification. , author=. Journal of Machine Learning Research , volume=

  41. [41]

    Journal of the American Statistical Association , volume=

    Nonparametric regression for spherical data , author=. Journal of the American Statistical Association , volume=. 2014 , publisher=

    2014

  42. [42]

    2009 , publisher=

    The elements of statistical learning: data mining, inference, and prediction , author=. 2009 , publisher=

    2009

  43. [43]

    Advances in Neural Information Processing Systems , pages=

    Global convergence of langevin dynamics based algorithms for nonconvex optimization , author=. Advances in Neural Information Processing Systems , pages=

  44. [44]

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

    The Neural Tangent Kernel in High Dimensions: Triple Descent and a Multi-Scale Theory of Generalization , author =. Proceedings of the 37th International Conference on Machine Learning , pages =. 2020 , editor =

    2020

  45. [45]

    Constructive Approximation , volume=

    On Early Stopping in Gradient Descent Learning , author=. Constructive Approximation , volume=

  46. [46]

    Communications on Pure and Applied Mathematics , volume=

    The generalization error of random features regression: Precise asymptotics and the double descent curve , author=. Communications on Pure and Applied Mathematics , volume=. 2022 , publisher=

    2022

  47. [47]

    Advances in neural information processing systems , volume=

    Random features for large-scale kernel machines , author=. Advances in neural information processing systems , volume=

  48. [48]

    Journal of Functional Analysis , volume=

    Generalization of an inequality by Talagrand and links with the logarithmic Sobolev inequality , author=. Journal of Functional Analysis , volume=. 2000 , publisher=

    2000

  49. [49]

    arXiv preprint arXiv:1910.11508 , year=

    Over Parameterized Two-level Neural Networks Can Learn Near Optimal Feature Representations , author=. arXiv preprint arXiv:1910.11508 , year=

    arXiv 1910

  50. [50]

    arXiv preprint arXiv:1904.04326 , year=

    A Comparative Analysis of the Optimization and Generalization Property of Two-layer Neural Network and Random Feature Models Under Gradient Descent Dynamics , author=. arXiv preprint arXiv:1904.04326 , year=

    arXiv 1904

  51. [51]

    the Thirty-Fourth AAAI Conference on Artificial Intelligence , year=

    Generalization Error Bounds of Gradient Descent for Learning Over-parameterized Deep ReLU Networks , author=. the Thirty-Fourth AAAI Conference on Artificial Intelligence , year=

  52. [52]

    Proceedings of the 34th International Conference on Machine Learning-Volume 70 , pages=

    Globally optimal gradient descent for a convnet with gaussian inputs , author=. Proceedings of the 34th International Conference on Machine Learning-Volume 70 , pages=. 2017 , organization=

    2017

  53. [53]

    International Conference on Machine Learning , pages=

    Overparameterized Nonlinear Learning: Gradient Descent Takes the Shortest Path? , author=. International Conference on Machine Learning , pages=

  54. [54]

    Training Over-parameterized Deep

    Zhang, Huishuai and Yu, Da and Chen, Wei and Liu, Tie-Yan , journal=. Training Over-parameterized Deep

  55. [55]

    arXiv preprint arXiv:1902.07111 , year=

    Global Convergence of Adaptive Gradient Methods for An Over-parameterized Neural Network , author=. arXiv preprint arXiv:1902.07111 , year=

    arXiv 1902

  56. [56]

    Advances in neural information processing systems , pages=

    Better mini-batch algorithms via accelerated gradient methods , author=. Advances in neural information processing systems , pages=

  57. [57]

    Zhurnal Vychislitel'noi Matematiki i Matematicheskoi Fiziki , volume=

    Gradient methods for minimizing functionals , author=. Zhurnal Vychislitel'noi Matematiki i Matematicheskoi Fiziki , volume=. 1963 , publisher=

    1963

  58. [58]

    Journal of Machine Learning Research , volume=

    Stochastic dual coordinate ascent methods for regularized loss minimization , author=. Journal of Machine Learning Research , volume=

  59. [59]

    Bell Labs Technical Journal , volume=

    The one-sided barrier problem for Gaussian noise , author=. Bell Labs Technical Journal , volume=. 1962 , publisher=

    1962

  60. [60]

    arXiv preprint arXiv:1312.6120 , year=

    Exact solutions to the nonlinear dynamics of learning in deep linear neural networks , author=. arXiv preprint arXiv:1312.6120 , year=

    Pith/arXiv arXiv

  61. [61]

    International Conference on Machine Learning , pages=

    Gradient descent with identity initialization efficiently learns positive definite linear transformations , author=. International Conference on Machine Learning , pages=

  62. [62]

    Electronic Communications in Probability , volume=

    A tail inequality for quadratic forms of subgaussian random vectors , author=. Electronic Communications in Probability , volume=. 2012 , publisher=

    2012

  63. [63]

    NIPS Tutorial , year=

    High-performance hardware for machine learning , author=. NIPS Tutorial , year=

  64. [64]

    Advances in neural information processing systems , pages=

    Sequence to sequence learning with neural networks , author=. Advances in neural information processing systems , pages=

  65. [65]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Going deeper with convolutions , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  66. [66]

    , author=

    Fast and Robust Neural Network Joint Models for Statistical Machine Translation. , author=. ACL (1) , pages=

  67. [67]

    arXiv preprint arXiv:1409.0473 , year=

    Neural machine translation by jointly learning to align and translate , author=. arXiv preprint arXiv:1409.0473 , year=

    Pith/arXiv arXiv

  68. [68]

    Neural networks , volume=

    Approximation capabilities of multilayer feedforward networks , author=. Neural networks , volume=. 1991 , publisher=

    1991

  69. [69]

    Advances In Neural Information Processing Systems , pages=

    Toward deeper understanding of neural networks: The power of initialization and a dual view on expressivity , author=. Advances In Neural Information Processing Systems , pages=

  70. [70]

    Conference on Learning Theory , pages=

    On the expressive power of deep learning: A tensor analysis , author=. Conference on Learning Theory , pages=

  71. [71]

    International Conference on Machine Learning , pages=

    Convolutional rectifier networks as generalized tensor decompositions , author=. International Conference on Machine Learning , pages=

  72. [72]

    arXiv preprint arXiv:1606.05336 , year=

    On the expressive power of deep neural networks , author=. arXiv preprint arXiv:1606.05336 , year=

    Pith/arXiv arXiv

  73. [73]

    Advances In Neural Information Processing Systems , pages=

    Exponential expressivity in deep neural networks through transient chaos , author=. Advances In Neural Information Processing Systems , pages=

  74. [74]

    Advances in neural information processing systems , pages=

    On the number of linear regions of deep neural networks , author=. Advances in neural information processing systems , pages=

  75. [75]

    Training , volume=

    Training a single sigmoidal neuron is hard , author=. Training , volume=. 2006 , publisher=

    2006

  76. [76]

    Advances in Neural Information Processing Systems , pages=

    On the computational efficiency of training neural networks , author=. Advances in Neural Information Processing Systems , pages=

  77. [77]

    arXiv preprint arXiv:1609.01037 , year=

    Distribution-specific hardness of learning neural networks , author=. arXiv preprint arXiv:1609.01037 , year=

    Pith/arXiv arXiv

  78. [78]

    International Conference on Machine Learning , pages=

    Failures of gradient-based deep learning , author=. International Conference on Machine Learning , pages=

  79. [79]

    arXiv preprint arXiv:1706.00687 , year=

    Weight Sharing is Crucial to Succesful Optimization , author=. arXiv preprint arXiv:1706.00687 , year=

    Pith/arXiv arXiv

  80. [80]

    Advances in neural information processing systems , pages=

    Training a 3-node neural network is NP-complete , author=. Advances in neural information processing systems , pages=

Showing first 80 references.