Pith. sign in

REVIEW 4 major objections 5 minor 33 references

Layer Separation Deep Learning Model with Auxiliary Variables for Partial Differential Equations

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that splitting a deep neural network into one-layer pieces via auxiliary variables turns the highly nonconvex PINN loss into a loss that is provably consistent with it and easier to minimize, yielding smaller residuals…

desk verdict Solid extension of the authors' LySep idea to PINNs with a proven elliptic consistency theorem and strong numerical gains, but the time-dependent proofs are skipped and the alternating algorithm lacks convergence guarantees. read the letter →

arxiv 2507.12766 v1 pith:YK5WL723 submitted 2025-07-17 cs.LG

Yaru Liu , Yiqi Gu This is my paper · ORCID
classification cs.LG
keywords physics-informedneuralnetworkslayerseparationauxiliaryvariablesalternatingdirectionoptimizationpartialdifferentialequationshigh-dimensionalPDEsconsistencyanalysisself-adaptiveweights
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

The paper tries to establish that the hardest part of training physics-informed neural networks for second-order linear PDEs, namely the deep, nonconvex coupling of layers inside the residual loss, can be removed without changing the optimization target. It introduces auxiliary variables for the output and derivatives of each layer, so the loss splits into terms that each couple at most two neighboring layers. New consistency theorems show that the layer-separated loss $J_S$ and the original PINN loss $J$ bound each other, so reducing $J_S$ is a faithful proxy for reducing the PDE residual. The accompanying alternating-direction algorithms update several variables in closed form, and the reported experiments on 2D/10D elliptic, 5D parabolic, and 5D hyperbolic equations show smaller final residual losses and smaller solution errors than vanilla-gradient-descent PINN. If correct, this gives a principled way to train deeper networks for PDEs without the usual vanishing-gradient stagnation.

What carries the argument

The central object is the self-adaptive weighted LySep loss $J_S$, built from auxiliary variables $a_l = W_l \sigma(a_{l-1}) + b_l \mathbf{1}^\top$ for layer outputs, $d_{li} = \partial_{x_i} a_l$ for first derivatives, and $q_i = W_2(\sigma''(a_1) * d_{1i} * d_{1i})$ for second-derivative combinations, coupled to the original network through quadratic penalty terms. The weights $\omega$ and diagonal matrices $D$ are chosen as functions of the network variables so that the penalty terms reproduce exactly the constant appearing in the one-sided inequality $J \le 2(d+1) C J_S$, and the choice $a_l = $ exact pre-activation, $d_{li} = $ exact derivative, $q_i = $ exact second-derivative combination makes $J_S \le J$. Because each squared term contains at most one nonlinear layer, the loss is convex in $W_3$, $b_1$, $b_2$, and $b_3$, and those variables are updated by solving least-squares linear systems, while the remaining variables are updated by gradient descent in an alternating Gauss-Seidel sweep.

What would settle it

Pick one of the four reported settings, say the 10-dimensional elliptic equation with a width-100 network, and rerun Algorithm 1 and the vanilla-gradient-descent PINN from the same ten random seeds for the same 2,000 iterations, recording $J_S$, $J_e$, and the $\ell^2$ test error at each checkpoint; if in any seed LySep's final $J_e$ or test error is not below PINN's, or if $J_S$ keeps decreasing while $J_e$ does not, the paper's practical claim fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that a deep PINN residual loss $J$ can be replaced by a layer-separated loss $J_S$ in which auxiliary variables $a_l$, $d_{li}$, and $q_i$ represent each layer's pre-activation output, its first derivatives, and its second-derivative combinations. The replacement is faithful: Theorem 2.1 (and its parabolic and hyperbolic analogues) states that $J \le 2(d+1) C J_S$ with a constant $C$ depending only on the activation's Lipschitz and boundedness constants, and Theorem 2.2 states that for any network weights there exist auxiliary variables making $J_S \le J$. Thus the two losses are consistent in both directions, and optimizing $J_S$ is a legitimate stand-in for optimizing $J$. The authors further claim, based on Tables 1 through 4, that their alternating-direction algorithms realize this proxy in practice, reaching final residual losses and $\ell^2$ solution errors far smaller than those of vanilla-gradient-descent PINN on the tested second-order linear PDEs, with the gap widening as networks get wider.

Load-bearing premise

The load-bearing premise is that the alternating-direction algorithms actually drive the layer-separated loss to a small value on the networks and PDEs tested; the paper proves consistency between the two losses but does not prove convergence of the algorithms.

Editorial extensions

If this is right

  • Optimizing a LySep loss is a valid stand-in for reducing the original PINN residual loss: any configuration with $J_S$ small is guaranteed by Theorem 2.1 to have $J$ small, and for every network weights there is an auxiliary configuration with $J_S$ no larger than $J$.
  • Because $J_S$ is shallow, closed-form updates for $W_3$ and the biases avoid backpropagating through the full network for those variables, which is intended to bypass the vanishing-gradient stagnation that slows vanilla PINN training.
  • The same construction is developed for elliptic, parabolic, and hyperbolic second-order equations, and the authors state that the strategy extends to higher-order and nonlinear PDEs at the cost of introducing more auxiliary variables.
  • On the tested examples, LySep's final residual loss and $\ell^2$ solution error decrease as network width increases, while vanilla PINN's errors stagnate or grow, and the numerical runs show $J_S$ and $J$ decreasing in parallel as the consistency theorems predict.
  • The observed parallel decrease of $J_S$ and $J$ across random seeds is direct numerical corroboration of the two-sided consistency bound, meaning the consistency inequality can be monitored during training as a sanity check.

Reading between the lines

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

  • Inference: Because every squared term in $J_S$ is shallow, the same decoupling should carry over to higher-order and nonlinear PDEs with more auxiliary variables; the authors say so in passing, and a concrete test would be a fourth-order biharmonic equation, where the second-derivative variables $q_i$ would multiply.
  • Inference: The consistency inequalities degrade with dimension, carrying factors $2(d+1)$, $2d+3$, and $2(d+2)$, so in very high dimensions the guaranteed bound is loose; a stress test at $d = 50$ to $d = 100$ would show whether the practical advantage persists or whether the constant is merely formal.
  • Inference: The reported algorithms still run gradient steps on $W_1$, $W_2$, and the auxiliary variables, so the closed-form updates alone are not the whole story; replacing those gradient steps with alternative stochastic optimizers inside the same LySep loss would test how much of the gain comes from the loss reformulation versus the update rules.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a layer separation (LySep) optimization framework for physics-informed neural networks (PINNs) solving second-order linear PDEs. Auxiliary variables represent the outputs and derivatives of each layer, decomposing the original deep loss into a sum of squared terms that couple only neighboring layers. For elliptic, parabolic, and hyperbolic equations the authors define LySep losses with self-adaptive weights and diagonal penalty matrices, prove (for the elliptic case) two consistency inequalities between the original PINN loss J_e and the LySep loss J_S, and give alternating-direction algorithms in which several variables have closed-form updates. Numerical experiments on 2D and 10D elliptic, 5D parabolic, and 5D hyperbolic problems with three-layer networks show that LySep achieves smaller final losses and smaller ℓ2 test errors than vanilla-gradient-descent PINN training.

Significance. If the consistency guarantees and the practical performance hold as stated, this is a useful contribution to optimization for physics-informed deep learning. The elliptic consistency theorems (Theorems 2.1 and 2.2) are proved with explicit constants in Appendix A, and the construction of self-adaptive weights and diagonal matrices is explicit rather than heuristic. The numerical study uses 10 random seeds and reports both the LySep loss and the original PINN loss, so the reader can directly verify the claimed simultaneous decrease. The main limitation is that the algorithmic side—the alternating-direction updates—lacks any convergence or descent guarantee, and the time-dependent consistency theorems are asserted without proof. These gaps make the paper's central claim ('optimizing the LySep loss indeed decreases the original mean square loss') an empirical observation rather than a fully established theorem.

major comments (4)
  1. [Section 3, Theorems 3.1–3.4] The consistency theorems for parabolic and hyperbolic equations are stated without proof. The text says 'The proof is similar to that of Theorem 2.1-2.2, so we do not present it here,' but these theorems are load-bearing for the paper's claim that the LySep model is consistent for time-dependent PDEs. Full proofs, or at least a detailed proof sketch that demonstrates the analogous inequalities with the stated constants, must be provided for the results used to interpret Tables 3 and 4.
  2. [Section 2.4, Algorithm 1 (and Algorithms 2–3)] The alternating-direction algorithm has no convergence guarantee. Theorems 2.1 and 2.2 are static statements about the loss functions at a single point; they do not imply that the iterates of Algorithm 1 reduce J_S or converge. Since the self-adaptive weights ω and matrices D depend on the current variables, and the gradient steps use a fixed learning rate τ with no Lipschitz-gradient or step-size condition, J_S is not even guaranteed to be nonincreasing across iterations. The conclusion in Section 5 that 'optimizing the LySep loss indeed decreases the original mean square loss' is therefore an algorithmic claim that needs either a convergence or monotonicity analysis, or an explicit statement restricting the consistency conclusion to the static loss landscape rather than the iterative process.
  3. [Section 4, Tables 1–4] All numerical experiments use three-layer networks (L=3), although the paper motivates LySep by the difficulties of deep networks and states that deeper formulations can be derived similarly. The central practical claim that LySep avoids bad local minima and gradient pathologies would be considerably strengthened by at least one experiment with L≥4. Without such evidence, the claims should be explicitly limited to the three-layer case, or the authors should provide a deeper-network formulation and test it.
  4. [Section 4, Tables 1–4 and Section 2.4] The empirical comparison conflates the LySep loss reformulation with the use of a different optimization algorithm: LySep is trained with alternating closed-form updates, while the PINN baseline is trained with vanilla gradient descent. To isolate the benefit of the LySep loss itself, the comparison should include a standard optimizer (e.g., Adam or L-BFGS) for the PINN baseline, or a full-gradient-descent solver for J_S. Without such a control, the reported accuracy improvements cannot be attributed solely to the LySep model rather than to the alternating-direction solver.
minor comments (5)
  1. [Section 2.4, Algorithms 2–3] Algorithms 2 and 3 refer to 'the linear system for b1 obtained in (3.14)' and (3.18), but the explicit linear systems are not displayed for the parabolic and hyperbolic cases. Please write out the exact update equations, as is done for the elliptic case in (2.20)–(2.22).
  2. [Section 4.1, Figure 4.1 caption] The phrase 'the the curves' contains a typo; it should read 'the curves'.
  3. [Section 4.1, Table 1] In Table 1 the LySep row reports three values under the column headed 'Actual loss J_e'. Please clarify the column headers so that J_S, J_e, and the test error are clearly distinguished.
  4. [Section 4.1.2, text before Table 2] The text says 'the mean and standard deviation of the loss J_e and training error', but the symbol \hat{E}_{\ell_2} is defined in (4.2) as a testing error. Please correct 'training error' to 'testing error'.
  5. [Section 2.3, definition of J_S] In (2.15), some penalty terms are written with the weighted norm \|\cdot\|_D and one term (the last, involving q_i) with the Frobenius norm. A brief remark explaining why the final term is unweighted would improve readability.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the LySep-PINN consistency theorems are derived inequalities, and the cited prior work [20] is not load-bearing.

full rationale

The paper's central claim is Theorem 2.1/2.2 and their time-dependent analogues: J_e <= 2(d+1) C J_S for all variables and, for every theta, existence of auxiliary variables with J_S <= J_e. Theorem 2.1 is proven in Appendix A by triangle inequalities and Lipschitz/boundedness assumptions on sigma; the self-adaptive weights omega and D are explicitly constructed so that the penalty terms dominate the error terms in the bound. This is a legitimate proof construction, not a circular one, because the inequality is derived from the definitions rather than assumed. Theorem 2.2 is the feasibility direction: setting a1 = W1X + b1 1^T, a2 = W2 sigma(a1) + b2 1^T, d1i = W1(:,i) 1^T, d2i = W2 (sigma'(a1)*d1i), and qi = W2 (sigma''(a1)*d1i*d1i) makes all penalty terms vanish; it is an existence statement, not a renaming of a fitted quantity. The only substantive reliance on prior work is a self-citation to [20] for the treatment of the R2 and T2i terms in the proof; the paper restates the main steps, and the PDE-specific loss decomposition and inequalities are new. No fitted parameter is relabeled as a prediction, no uniqueness theorem is imported from the authors' prior work, and the numerical comparisons are against a vanilla PINN baseline rather than being forced by the construction. The main weakness is that Algorithms 1-3 lack a convergence or stationarity proof, so the static consistency inequalities do not by themselves guarantee that the alternating updates drive J_S down; this is an unsupported practical claim, not a circular derivation, and it does not affect the consistency theorems. Overall, the derivation chain is self-contained apart from one minor, non-load-bearing self-citation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central derivation uses no fitted physical constants and invents no physical entities. The auxiliary variables a_l, d_li, and q_i are optimization variables with definitional constraints, not new physical content. The load-bearing assumptions are smoothness of the activation, regularity of the PDE solution, and unproved convergence of the alternating algorithm.

free parameters (3)
  • learning rate tau = not specified
    Used in gradient descent updates for W1, W2, a1, a2, d1i, d2i, and qi in Algorithms 1-3; the paper never gives its value or schedule, so numerical results depend on an unreported choice.
  • iteration count N_k = 2000 in all experiments
    Stopping budget for Algorithms 1-3; reported losses and errors are final values at this budget.
  • network width M = 30, 50, 80, 100 for 2D; 20, 50, 100, 150 for 10D; 20, 50, 80, 100 for parabolic and hyperbolic
    Chosen network capacity; the claim that LySep improves with width is specific to these sizes.
assumptions (4)
  • domain assumption Activation sigma and its first two derivatives are Lipschitz continuous and bounded
    Assumed in Theorems 2.1, 3.1, and 3.3; the experiments use sin activation, which satisfies the assumption, but ReLU does not.
  • domain assumption Solutions of the PDEs exist and are regular enough for strong-form pointwise residual training
    The PINN formulation evaluates D[u,x] pointwise; the paper does not state regularity conditions on u, f, and c beyond c(x) > delta.
  • ad hoc to paper The alternating direction algorithm converges to a low-value minimizer of J_S
    No convergence or stationarity theorem is given; the consistency inequalities only relate J_e and J_S, not the algorithm's trajectory.
  • domain assumption DNNs of the used widths can approximate the PDE solution well enough for the residual loss to be a faithful proxy
    The paper cites Barron-type approximation results but does not verify approximation error for the specific networks and PDEs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Layer Separation Deep Learning Model with Auxiliary Variables for Partial Differential Equations." pith.science (2026). https://pith.science/paper/YK5WL723

@misc{pith2026250712766,
  author       = {Pith},
  title        = {Pith review of: Layer Separation Deep Learning Model with Auxiliary Variables for Partial Differential Equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YK5WL723}},
  note         = {Machine review of arXiv:2507.12766}
}
read the original abstract

In this paper, we propose a new optimization framework, the layer separation (LySep) model, to improve the deep learning-based methods in solving partial differential equations. Due to the highly non-convex nature of the loss function in deep learning, existing optimization algorithms often converge to suboptimal local minima or suffer from gradient explosion or vanishing, resulting in poor performance. To address these issues, we introduce auxiliary variables to separate the layers of deep neural networks. Specifically, the output and its derivatives of each layer are represented by auxiliary variables, effectively decomposing the deep architecture into a series of shallow architectures. New loss functions with auxiliary variables are established, in which only variables from two neighboring layers are coupled. Corresponding algorithms based on alternating directions are developed, where many variables can be updated optimally in closed forms. Moreover, we provide theoretical analyses demonstrating the consistency between the LySep model and the original deep model. High-dimensional numerical results validate our theory and demonstrate the advantages of LySep in minimizing loss and reducing solution error.

Figures

Figures reproduced from arXiv: 2507.12766 by the authors.

Figure 4.1
Figure 4.1. The loss J e versus iterations in solving the 2-D elliptic example with the PINN and LySep models (10 tests). in theory. Our numerical results show that the J e ≤ 2(d + 1) · C · J e S relation in the LySep model always holds under all seed conditions in high-dimensional case. (L, M) Models Actual loss J e Eˆ ℓ2 (3, 20) PINN 2.82e-06±1.93e-06 7.29e-03±2.50e-03 LySep 1.09e-07±7.98e-08 1.25e-07± 9.67e-08 1.12e-03±5.51e… view at source ↗
Figure 4.2
Figure 4.2. The losses J e S and J e versus iterations in solving the 2-D elliptic example with the LySep model (10 tests). 4.2. Parabolic Problem In the third example, we use the PINN and LySep models to solve the parabolic problem (3.1) with c(x) = 1 d Pd i=1 xi + 2 and let the true solution be u(t, x) =  exp  − t d  − 1  sin 1 d X d i=1 x 2 i − 1 X d i=1 cos  xi √ d  , where d = 5. We set Ω = {x ∈ R 5 , kxk2 ≤ 1} a… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 31 canonical work pages

  1. [1]

    Yserentant, Sparse grid spaces for the numerical solu tion of the electronic schr¨ odinger equation, Numer

    H. Yserentant, Sparse grid spaces for the numerical solu tion of the electronic schr¨ odinger equation, Numer. Math. 101 (2005) 381–389

    work page 2005

  2. [2]

    T. Lee, F. W ang, R. Newell, Robust model-order reduction of complex biological processes, J. Process Control 12 (200 2) 807–821

    work page

  3. [3]

    Ehrhardt, R

    M. Ehrhardt, R. E. R. E Mickens, A fast, stable and accurat e numerical method for the Black–Scholes equation of American options, Int. J. Theor. Appl. Finance 11 (2008) 471 –501

    work page 2008

  4. [4]

    S. Cai, Z. Mao, Z. W ang, M. Yin, G. E. Karniadakis, Physics -informed neural networks (PINNs) for fluid mechanics: A review, Acta Mech. Sin. (2021) 1727–1738

    work page 2021

  5. [5]

    S. Cai, Z. W ang, S. W ang, P. Perdikaris, G. E. Karniadakis , Physics-informed neural networks for heat transfer probl ems, J. Heat Transf. 143 (2021) 060801

    work page 2021

  6. [6]

    A. D. Jagtap, G. E. Karniadakis, Extended physics-infor med neural networks (XPINNs): A generalized space-time dom ain decomposition based deep learning framework for nonlinear partial differential equations, Commun. Comput. Phys. 28 (2020)

    work page 2020

  7. [7]

    Z. Mao, A. D. Jagtap, G. E. Karniadakis, Physics-informe d neural networks for high-speed flows, Comput. Methods Appl . Mech. Engrg. 360 (2020) 112789

    work page 2020

  8. [8]

    G. Pang, L. Lu, G. E. Karniadakis, fPINNs: Fractional phy sics-informed neural networks, SIAM J. Sci. Comput. 41 (2019) A2603–A2626

    work page 2019

Show all 33 references
  1. [9]

    Raissi, P

    M. Raissi, P. Perdikaris, G. E. Karniadakis, Physics-in formed neural networks: A deep learning framework for solvi ng forward and inverse problems involving nonlinear partial d ifferential equations, J. Comput. Phys. 378 (2019) 686–707

    2019

  2. [10]

    C. Rao, H. Sun, Y. Liu, Physics-informed deep learning f or incompressible laminar flows, Theor. Appl. Mech. Lett. 10 (2020) 207–212

    2020

  3. [11]

    Yu, The deep Ritz method: a deep learning-based numer ical algorithm for solving variational problems, Commun

    B. Yu, The deep Ritz method: a deep learning-based numer ical algorithm for solving variational problems, Commun. Math. Stat. 6 (2018) 1–12

    2018

  4. [12]

    Y. Gu, H. Yang, C. Zhou, Selectnet: Self-paced learning for high-dimensional partial differential equations, J. Co mput. Phys. 441 (2021) 110444

    2021

  5. [13]

    Y. Zang, G. Bao, X. Ye, H. Zhou, W eak adversarial network s for high-dimensional partial differential equations, J. Comput. Phys. 411 (2020) 109409

    2020

  6. [14]

    G. Bao, D. W ang, B. Zou, W ANCO: weak adversarial network s for constrained optimization problems, 2024. Https://arxiv.org/abs/2407.03647

    2024 arXiv

  7. [15]

    Y. Gao, Y. Gu, M. Ng, Gradient descent finds the global opt ima of two-layer physics-informed neural networks, in: Proceedings of the 40th International Conference on Machin e Learning, 2023, pp. 10676–10707

    2023

  8. [16]

    T. Luo, H. Yang, Two-layer neural networks for partial d ifferential equations: Optimization and generalization th eory,

  9. [17]

    Y. Gu, C. W ang, H. Yang, Structure probing neural networ k deflation, J. Comput. Phys. 434 (2021) 110231

    2021

  10. [18]

    R¨ ognvaldsson, On langevin updating in multilayer p erceptrons, Neural Comput

    T. R¨ ognvaldsson, On langevin updating in multilayer p erceptrons, Neural Comput. 6 (1994) 916–926

    1994

  11. [19]

    Erhan, P

    D. Erhan, P. A. Manzagol, Y. Bengio, S. Bengio, P. Vincen t, The difficulty of training deep architectures and the effect of unsupervised pre-training, in: Proceedings of the 12th Int ernational Conference on Artificial Intelligence and Stati stics, 2009, pp. 153–160

    2009

  12. [20]

    Y. Liu, Y. Gu, M. K. Ng, Deep learning optimization using self-adaptive weighted auxiliary variables, 2025. Https://arxiv.org/abs/2504.21501

    2025 arXiv

  13. [21]

    Zhang, Z

    S. Zhang, Z. Shen, H. Yang, Deep Network Approximation: Achieving Arbitrary Accuracy with Fixed Number of Neurons, J. Mach. Learn. Res. 23 (2022) 1–60

    2022

  14. [22]

    Y. Jiao, Y. Lai, X. Lu, F. W ang, J. Z. Yang, Y. Yang, Deep ne ural networks with relu-sine-exponential activations bre ak curse of dimensionality in approximation on h¨ older class, SIAM J. Math. Anal. 55 (2023) 3635–3649

    2023

  15. [23]

    Z. Shen, H. Yang, S. Zhang, Neural network approximatio n: Three hidden layers are enough, Neural Netw. 141 (2021) 160–173

    2021

  16. [24]

    Z. Shen, H. Yang, S. Zhang, Deep network with approximat ion error being reciprocal of width to power of square root of depth, Neural Comput. 33 (2021) 1005–1036. 23

    2021

  17. [25]

    A. R. Barron, Neural net approximation, in: Proceeding s of the 7th Yale W orkshop on Adaptive and Learning Systems, 1992, pp. 69–72

    1992

  18. [26]

    A. R. Barron, Universal approximation bounds for super positions of a sigmoidal function, IEEE Trans. Inform. Theo ry 39 (1993) 930–945

    1993

  19. [27]

    Caragea, P

    A. Caragea, P. Petersen, F. Voigtlaender, Neural netwo rk approximation and estimation of classifiers with classifi cation boundary in a Barron class, Ann. Appl. Probab. 33 (2023) 3039 –3079

    2023

  20. [28]

    W. E, C. Ma, S. W ojtowytsch, L. W u, Towards a mathematica l understanding of neural network-based machine learning: what we know and what we don’t, 2020. Https://arxiv.org/abs /2009.10713

    2020 arXiv

  21. [29]

    W. E, C. Ma, L. W u, The Barron space and the flow-induced fu nction spaces for neural network models, Constr. Approx. 55 (2022) 369–406

    2022

  22. [30]

    J. M. Klusowski, A. R. Barron, Approximation by combina tions of ReLU and squared ReLU ridge functions with ℓ1 and ℓ0 controls, IEEE Trans. Inform. Theory 64 (2018) 7649–7656

    2018

  23. [31]

    J. W. Siegel, J. Xu, Approximation rates for neural netw orks with general activation functions, Neural Netw. 128 (2 020) 313–321

  24. [32]

    J. W. Siegel, J. Xu, High-order approximation rates for shallow neural networks with cosine and ReLU k activation functions, Appl. Comput. Harmon. Anal. 58 (2022) 1–26. 24

    2022

  25. [2020]

    Https://arxiv.org/abs/2006.15733

    2006 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.