arxiv: 2605.07286 · v1 · submitted 2026-05-08 · 🧮 math.NA · cs.LG· cs.NA· physics.comp-ph

Recognition: no theorem link

Sparse Random-Feature Neural Networks with Krylov-Based SVD for Singularly Perturbed ODE

Kevin Kurian Thomas Vaidyan, Siddharth Rout

Pith reviewed 2026-05-11 01:11 UTC · model grok-4.3

classification 🧮 math.NA cs.LGcs.NAphysics.comp-ph

keywords random-feature neural networkssparse singular value decompositionsingularly perturbed ODEsconvection-diffusion equationsKrylov methodsnumerical conditioningleast squares problemsscientific machine learning

0 comments

The pith

Sparse random-feature networks with structured sparsity and sSVD maintain accuracy on stiff convection-diffusion equations while improving efficiency.

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

This paper aims to show that random-feature neural networks can be made more scalable and stable for singularly perturbed ODEs by adding structured sparsity to the hidden layer activations and solving the output weights with sparse SVD. The activation matrix in standard RFNNs tends to be low-rank and ill-conditioned, limiting their use on stiff problems like convection-diffusion equations with strong advection. By increasing the rank through sparsity and using a Krylov-based sparse SVD method, the approach achieves better training efficiency and robustness without sacrificing solution accuracy on benchmark one-dimensional cases. Readers interested in numerical methods for stiff systems would find this relevant as it provides a practical way to handle computational challenges in these networks.

Core claim

The central claim is that integrating structured sparsity into the hidden layer activations of RFNNs increases the rank of the activation matrix and allows the use of sparse singular value decomposition via Lanczos-Golub-Kahan bidiagonalization to efficiently solve the ill-conditioned least squares problem for output weights, resulting in accurate solutions to singularly perturbed ODEs with substantial improvements in efficiency and robustness over dense RFNN implementations.

What carries the argument

Structured sparsity in hidden layer activations paired with Krylov-based sparse SVD (sSVD) using Lanczos-Golub-Kahan bidiagonalization to handle the low-rank and ill-conditioned activation matrix.

If this is right

The proposed method maintains or improves accuracy for 1D convection-diffusion equations with stronger advection.
It achieves substantial gains in training efficiency compared to dense RFNNs.
Robustness is improved due to better conditioning and scalability of the solver.
The framework caters to high-dimensional or stiff systems by addressing numerical stability issues.

Where Pith is reading between the lines

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

This sparsity technique might be adaptable to other machine learning models facing similar low-rank activation issues in scientific computing.
The requirement for an orthogonalization step in the sSVD process highlights a potential area for further optimization in iterative solvers for neural network training.
If effective here, the method could influence hybrid approaches combining neural networks with traditional numerical methods for singularly perturbed problems.

Load-bearing premise

The load-bearing premise is that the structured sparsity sufficiently raises the rank and improves the conditioning of the activation matrix so that sparse SVD can deliver stable and accurate least squares solutions for these problems.

What would settle it

Observing whether the sparse RFNN with sSVD produces solution errors comparable to or lower than dense RFNNs on the benchmark equations as advection strength increases, while also showing reduced training times, would test the claim; failure to do so would falsify it.

Figures

Figures reproduced from arXiv: 2605.07286 by Kevin Kurian Thomas Vaidyan, Siddharth Rout.

**Figure 1.** Figure 1: True solution for P e = −1e4. 5 Experiments 5.1 Understanding the problem: Solving Non-smooth ODE (Steady 1D Convection-Diffusion) Governing Equation: u dϕ dx = D d 2ϕ dx2 (30) Boundary Conditions: ϕ(0) = ϕ0, ϕ(L) = ϕL (31) Exact Solution: ϕ(x) = ϕ0 + (ϕL − ϕ0) e u D x − 1 e u D L − 1 (32) In this work, we take L = 1, ϕ0 = 0, ϕL = 1 and P e = u D = −1e3. Key Behavior: • P e ≫ 1: Convection-dominated → Ultr… view at source ↗

**Figure 2.** Figure 2: Visualization of activation matrices. University of British Columbia [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

**Figure 3.** Figure 3: Singular values for Sparse Hard random matrix [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

**Figure 4.** Figure 4: Degenerate singular values for Sparse Hard random matrix [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗

**Figure 5.** Figure 5: Visualization of supposed to be orthonormal matrices. [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

**Figure 6.** Figure 6: Singular values for Sparse Hard random matrix [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗

**Figure 7.** Figure 7: Singular values obtained using sparse SVD using LGK with and without full-orthogonalization. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗

**Figure 8.** Figure 8: Singular values obtained using sparse SVD using LGK with full-orthogonalization. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗

**Figure 9.** Figure 9: Spy on activation matrices. (a) Kernel width 1e − 5. (b) Kernel width 1e − 6, [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗

**Figure 10.** Figure 10: Our solution of 1D steady convection-diffusion equation for [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗

read the original abstract

Random-feature neural networks (RFNNs), including architectures with fixed hidden layers and analytically determined output weights, offer fast training but often suffer from issues due to dense representations of the hidden layer activation. Their reliance on dense feature mappings and least squares solvers can limit scalability and numerical stability, particularly for high-dimensional or stiff systems. Specifically, the activation matrix is observed to be low-rank and extremely ill-conditioned. In this work, we propose a sparse framework for RFNNs that integrates structured sparsity into the hidden layer activations that increases the rank and employs Sparse Singular Value Decomposition (sSVD) for solving the resulting linear least squares problem scalably and efficiently while catering to the bad condition number. We explore the theory behind Lanczos-Golub-Kahan Bidiagonalization technique for sparse SVD and conduct some experiments to identify some limitations and justify the requirement for orthogonalization step in our application. Then, we demonstrate that the proposed method maintains or improves solution accuracy for solving the benchmark one-dimensional steady convection-diffusion equations case having stronger advection, while achieving substantial gains in training efficiency and robustness compared to standard dense implementations.

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.

Desk Editor's Note private letter to a colleague

The paper introduces structured sparsity plus Krylov sSVD in random-feature networks for 1D singularly perturbed ODEs, but supplies no matrix diagnostics to show the sparsity actually fixes the claimed ill-conditioning.

read the letter

The main thing here is a sparse random-feature neural network that imposes structured sparsity on the hidden-layer activations and then uses a Krylov-based sparse SVD to compute the output weights. This is aimed at solving one-dimensional singularly perturbed convection-diffusion equations where the activation matrix is low-rank and ill-conditioned. They show that this combination maintains or improves accuracy on benchmark cases with stronger advection while cutting training time and improving robustness over dense RFNNs. The theory section on Lanczos-Golub-Kahan bidiagonalization and the need for an extra orthogonalization step is a reasonable addition to handle the conditioning issues. The approach is new in tying structured sparsity directly to the rank and conditioning problem in this RFNN setting for ODEs. The experiments appear to support the efficiency claims, at least at the level described. The soft spot is that the central mechanism—how the sparsity pattern actually raises the numerical rank and lowers the condition number—is not backed by the kind of matrix diagnostics one would want. There are no reported condition-number values, rank estimates, or singular-value plots comparing the dense and sparse activation matrices on the test problems. Without that, it's difficult to confirm that the sparsity is addressing the ill-conditioning the abstract identifies as the main obstacle, rather than the orthogonalization step or other factors doing the work. This paper is for researchers in numerical analysis who are exploring neural-network-based solvers for stiff or singularly perturbed problems. It could be useful to someone looking for ways to scale RFNNs without full dense linear algebra. I would recommend sending it to peer review. The idea is concrete and the application is well-defined, so referees can check the missing diagnostics and the experimental details.

Referee Report

1 major / 1 minor

Summary. The manuscript proposes a sparse random-feature neural network (RFNN) framework that imposes structured sparsity on hidden-layer activations to raise numerical rank, then solves the resulting output-weight least-squares problem via a Krylov-based sparse SVD (Lanczos-Golub-Kahan bidiagonalization plus orthogonalization). The central claim is that this combination maintains or improves accuracy on 1D steady convection-diffusion benchmarks with strong advection (small ε) while delivering substantial gains in training efficiency and robustness relative to dense RFNN implementations.

Significance. If the sparsity pattern demonstrably improves conditioning and rank, the approach would supply a scalable, sparse-linear-algebra route to stable RFNN solutions for singularly perturbed problems. The work explicitly explores the limitations of plain sSVD and the necessity of the orthogonalization step, which is a constructive contribution; however, the absence of direct matrix-property diagnostics limits the immediate impact.

major comments (1)

[Abstract and numerical-experiments section] Abstract and the numerical-experiments section: the assertion that structured sparsity 'increases the rank and caters to the bad condition number' of the activation matrix is load-bearing for the claim that sSVD yields stable, accurate solutions at small ε. No condition-number tables, numerical-rank counts, or singular-value spectra comparing the dense and sparse activation matrices on the reported convection-diffusion benchmarks are supplied, leaving the weakest assumption unquantified.

minor comments (1)

The phrase 'some experiments to identify some limitations' in the abstract is vague; a concise enumeration of the observed limitations and the precise role of the orthogonalization step would improve readability.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation of major revision. The central concern is the lack of quantitative diagnostics supporting the claim that structured sparsity improves the rank and conditioning of the activation matrix. We address this point directly below and will revise the manuscript to include the requested evidence.

read point-by-point responses

Referee: Abstract and the numerical-experiments section: the assertion that structured sparsity 'increases the rank and caters to the bad condition number' of the activation matrix is load-bearing for the claim that sSVD yields stable, accurate solutions at small ε. No condition-number tables, numerical-rank counts, or singular-value spectra comparing the dense and sparse activation matrices on the reported convection-diffusion benchmarks are supplied, leaving the weakest assumption unquantified.

Authors: We agree that the current manuscript does not supply direct matrix-property diagnostics such as condition-number tables, numerical-rank counts, or singular-value spectra comparing dense versus sparse activation matrices on the convection-diffusion benchmarks. This leaves the key assumption unquantified and weakens the support for the stability claims at small ε. In the revised version we will add these diagnostics to the numerical-experiments section: tables reporting 2-norm condition numbers and effective numerical ranks (e.g., number of singular values above a tolerance) for both dense and sparse cases across the reported ε values; and plots or tabulated singular-value spectra that illustrate the rank increase and conditioning improvement induced by the structured sparsity. These additions will make the benefit of the sparsity pattern explicit and directly address the referee's concern. revision: yes

Circularity Check

0 steps flagged

No circularity; new sparsity pattern and sSVD integration are independent of inputs

full rationale

The paper proposes a sparse RFNN framework by adding structured sparsity to hidden activations and using Lanczos-Golub-Kahan bidiagonalization for sSVD on the resulting least-squares problem. This builds on standard RFNN and numerical linear algebra techniques without any self-definitional loops, fitted parameters renamed as predictions, or load-bearing self-citations. The claim that sparsity increases rank and improves conditioning is presented as a design choice justified by experiments, not derived by construction from the target accuracy metrics. No uniqueness theorems or ansatzes are smuggled via self-citation. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The paper relies on standard assumptions about RFNN properties and introduces sparsity as a modification; no new entities postulated. Free parameters include typical NN hyperparameters and the specific sparsity structure.

free parameters (2)

sparsity level or structure parameter
The degree of structured sparsity introduced in hidden layer activations is likely chosen or tuned, affecting rank and conditioning.
number of random features
Standard in RFNNs, the number of features is a hyperparameter that impacts the model.

axioms (2)

domain assumption The activation matrix in RFNNs is low-rank and ill-conditioned for dense cases
Stated in abstract as observed.
standard math Lanczos-Golub-Kahan bidiagonalization can be used for sparse SVD with orthogonalization for stability
Explored in theory section per abstract.

pith-pipeline@v0.9.0 · 5507 in / 1597 out tokens · 64623 ms · 2026-05-11T01:11:11.696361+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

46 extracted references · 46 canonical work pages · 2 internal anchors

[1]

Speech recognition using deep neural networks: A systematic review.IEEE Access, 7:19143–19165, 2019

Ali Bou Nassif, Ismail Shahin, Imtinan Attili, Mohammad Azzeh, and Khaled Shaalan. Speech recognition using deep neural networks: A systematic review.IEEE Access, 7:19143–19165, 2019

work page 2019
[2]

A survey of convolutional neural networks: Analysis, applications, and prospects.IEEE Transactions on Neural Networks and Learning Systems, 33(12):6999– 7019, 2022

Zewen Li, Fan Liu, Wenjie Yang, Shouheng Peng, and Jun Zhou. A survey of convolutional neural networks: Analysis, applications, and prospects.IEEE Transactions on Neural Networks and Learning Systems, 33(12):6999– 7019, 2022

work page 2022
[3]

Scalable diffusion models with transformers

William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195–4205, 2023

work page 2023
[4]

A survey on evaluation of large language models.ACM Transactions on Intelligent Systems and Technology, 15(3):1–45, 2024

Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models.ACM Transactions on Intelligent Systems and Technology, 15(3):1–45, 2024

work page 2024
[5]

Large language models in medicine.Nature medicine, 29(8):1930–1940, 2023

Arun James Thirunavukarasu, Darren Shu Jeng Ting, Kabilan Elangovan, Laura Gutierrez, Ting Fang Tan, and Daniel Shu Wei Ting. Large language models in medicine.Nature medicine, 29(8):1930–1940, 2023

work page 1930
[6]

Scientific machine learning benchmarks

Jeyan Thiyagalingam, Mallikarjun Shankar, Geoffrey Fox, and Tony Hey. Scientific machine learning benchmarks. Nature Reviews Physics, 4(6):413–420, 2022

work page 2022
[7]

Scientific machine learning through physics–informed neural networks: Where we are and what’s next

Salvatore Cuomo, Vincenzo Schiano Di Cola, Fabio Giampaolo, Gianluigi Rozza, Maziar Raissi, and Francesco Piccialli. Scientific machine learning through physics–informed neural networks: Where we are and what’s next. Journal of Scientific Computing, 92(3):88, 2022

work page 2022
[8]

FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators

Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopadhyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, Kamyar Azizzadenesheli, et al. Fourcastnet: A global data-driven high- resolution weather model using adaptive fourier neural operators.arXiv preprint arXiv:2202.11214, 2022

work page internal anchor Pith review arXiv 2022
[9]

KAN: Kolmogorov-Arnold Networks

Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljaˇci´c, Thomas Y Hou, and Max Tegmark. Kan: Kolmogorov-arnold networks.arXiv preprint arXiv:2404.19756, 2024

work page internal anchor Pith review arXiv 2024
[10]

Extreme learning machine: theory and applications

Guang-Bin Huang, Qin-Yu Zhu, and Chee-Kheong Siew. Extreme learning machine: theory and applications. Neurocomputing, 70(1-3):489–501, 2006. University of British Columbia

work page 2006
[11]

Six lectures on linearized neural networks.arXiv preprint arXiv:2308.13431, 2023

Theodor Misiakiewicz and Andrea Montanari. Six lectures on linearized neural networks.arXiv preprint arXiv:2308.13431, 2023

work page arXiv 2023
[12]

Universal approximation using incremental constructive feedforward networks with random hidden nodes.IEEE Transactions on Neural Networks, 17:879–892, 2006

Guangbin Huang, Lei Chen, and Chee Kheong Siew. Universal approximation using incremental constructive feedforward networks with random hidden nodes.IEEE Transactions on Neural Networks, 17:879–892, 2006

work page 2006
[13]

Convex incremental extreme learning machine.Neurocomputing, 70(16):3056– 3062, 2007

Guang-Bin Huang and Lei Chen. Convex incremental extreme learning machine.Neurocomputing, 70(16):3056– 3062, 2007

work page 2007
[14]

MIT press, 2024

Francis Bach.Learning theory from first principles. MIT press, 2024

work page 2024
[15]

Harnessing nonlinearity: Predicting chaotic systems and saving energy in wireless communication.science, 304(5667):78–80, 2004

Herbert Jaeger and Harald Haas. Harnessing nonlinearity: Predicting chaotic systems and saving energy in wireless communication.science, 304(5667):78–80, 2004

work page 2004
[16]

Random features for large-scale kernel machines.Advances in neural information processing systems, 20, 2007

Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines.Advances in neural information processing systems, 20, 2007

work page 2007
[17]

Generalization properties of learning with random features.Advances in neural information processing systems, 30, 2017

Alessandro Rudi and Lorenzo Rosasco. Generalization properties of learning with random features.Advances in neural information processing systems, 30, 2017

work page 2017
[18]

The generalization error of random features regression: Precise asymptotics and the double descent curve.Communications on Pure and Applied Mathematics, 75(4):667–766, 2022

Song Mei and Andrea Montanari. The generalization error of random features regression: Precise asymptotics and the double descent curve.Communications on Pure and Applied Mathematics, 75(4):667–766, 2022

work page 2022
[19]

Physics informed extreme learning machine (pielm)–a rapid method for the numerical solution of partial differential equations.Neurocomputing, 391:96–118, 2020

Vikas Dwivedi and Balaji Srinivasan. Physics informed extreme learning machine (pielm)–a rapid method for the numerical solution of partial differential equations.Neurocomputing, 391:96–118, 2020

work page 2020
[20]

Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational physics, 378:686–707, 2019

work page 2019
[21]

Physics- informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021

George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics- informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021

work page 2021
[22]

Understanding and mitigating gradient flow pathologies in physics-informed neural networks.SIAM Journal on Scientific Computing, 43(5):A3055–A3081, 2021

Sifan Wang, Yujun Teng, and Paris Perdikaris. Understanding and mitigating gradient flow pathologies in physics-informed neural networks.SIAM Journal on Scientific Computing, 43(5):A3055–A3081, 2021

work page 2021
[23]

Near-optimal sketchy natural gradients for physics-informed neural networks

Maricela Best Mckay, Avleen Kaur, Chen Greif, and Brian Wetton. Near-optimal sketchy natural gradients for physics-informed neural networks. InForty-second International Conference on Machine Learning

work page
[24]

An analysis and solution of ill-conditioning in physics-informed neural networks

Wenbo Cao and Weiwei Zhang. An analysis and solution of ill-conditioning in physics-informed neural networks. Journal of Computational Physics, 520:113494, 2025

work page 2025
[25]

Numerical approximation in cfd problems using physics informed machine learning.Indian Institute of Technology Madras, 2019

Siddharth Rout. Numerical approximation in cfd problems using physics informed machine learning.Indian Institute of Technology Madras, 2019

work page 2019
[26]

Extended physics-informed neural networks (xpinns): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equations

Ameya D Jagtap and George Em Karniadakis. Extended physics-informed neural networks (xpinns): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equations. Communications in Computational Physics, 28(5), 2020

work page 2020
[27]

Enhancing training of physics- informed neural networks using domain decomposition–based preconditioning strategies.SIAM Journal on Scientific Computing, 46(5):S46–S67, 2024

Alena Kopaniˇcáková, Hardik Kothari, George E Karniadakis, and Rolf Krause. Enhancing training of physics- informed neural networks using domain decomposition–based preconditioning strategies.SIAM Journal on Scientific Computing, 46(5):S46–S67, 2024

work page 2024
[28]

Preconditioning for physics-informed neural networks.arXiv preprint arXiv:2402.00531, 2024

Songming Liu, Chang Su, Jiachen Yao, Zhongkai Hao, Hang Su, Youjia Wu, and Jun Zhu. Preconditioning for physics-informed neural networks.arXiv preprint arXiv:2402.00531, 2024

work page arXiv 2024
[29]

A matrix preconditioning framework for physics-informed neural networks based on the adjoint method.Physics of Fluids, 37(9), 2025

Jiahao Song, Wenbo Cao, and Weiwei Zhang. A matrix preconditioning framework for physics-informed neural networks based on the adjoint method.Physics of Fluids, 37(9), 2025

work page 2025
[30]

A preconditioned quasi-newton optimizer for efficient training of pinns

Shahbaz Ahmad and Muhammad Israr. A preconditioned quasi-newton optimizer for efficient training of pinns. Machine Learning for Computational Science and Engineering, 1(2):34, 2025

work page 2025
[31]

Multi-stage neural networks: Function approximator of machine precision

Yongji Wang and Ching-Yao Lai. Multi-stage neural networks: Function approximator of machine precision. Journal of Computational Physics, 504:112865, 2024

work page 2024
[32]

Physics-guided multistage neural network: A physically guided network for step initial values and dispersive shock wave phenomena.Physical Review E, 110(6):065307, 2024

Wen-Xuan Yuan and Rui Guo. Physics-guided multistage neural network: A physically guided network for step initial values and dispersive shock wave phenomena.Physical Review E, 110(6):065307, 2024

work page 2024
[33]

Multistage physics informed neural network for solving coupled multiphysics problems in material degradation and fluid dynamics.Engineering with Computers, pages 1–31, 2025

Mahmoud Khadijeh, Veronica Cerqueglini, Cor Kasbergen, Sandra Erkens, and Aikaterini Varveri. Multistage physics informed neural network for solving coupled multiphysics problems in material degradation and fluid dynamics.Engineering with Computers, pages 1–31, 2025

work page 2025
[34]

Physics-informed extreme learning machine (pielm) for stefan problems.Computer Methods in Applied Mechanics and Engineering, 441:118015, 2025

Fei Ren, Pei-Zhi Zhuang, Xiaohui Chen, Hai-Sui Yu, and He Yang. Physics-informed extreme learning machine (pielm) for stefan problems.Computer Methods in Applied Mechanics and Engineering, 441:118015, 2025. University of British Columbia

work page 2025
[35]

Physics-informed extreme learning machine framework for solving linear elasticity mechanics problems.International Journal of Solids and Structures, 309:113157, 2025

Qimin Wang, Chao Li, Sheng Zhang, Chen Zhou, and Yanping Zhou. Physics-informed extreme learning machine framework for solving linear elasticity mechanics problems.International Journal of Solids and Structures, 309:113157, 2025

work page 2025
[36]

Eig-pielm: A mesh- free approach for efficient eigen-analysis with physics-informed extreme learning machines.arXiv preprint arXiv:2508.15343, 2025

Rishi Mishra, Ganapathy Krishnamurthi, Balaji Srinivasan, Sundararajan Natarajan, et al. Eig-pielm: A mesh- free approach for efficient eigen-analysis with physics-informed extreme learning machines.arXiv preprint arXiv:2508.15343, 2025

work page arXiv 2025
[37]

Li Huang, Liang Chen, and Rongchuan Bai. Physics-informed extreme learning machine applied for eigenmode analysis of waveguides and transmission lines.International Journal of RF and Microwave Computer-Aided Engineering, 2025(1):6233356, 2025

work page 2025
[38]

Fast, convex and conditioned single-layer network for learning multi-fidelity univariate data and linear differential equations

Siddharth Rout. Fast, convex and conditioned single-layer network for learning multi-fidelity univariate data and linear differential equations. InAI&PDE: ICLR 2026 Workshop on AI and Partial Differential Equations

work page 2026
[39]

Selm: Semi-supervised elm with application in sparse calibrated location estimation.Neurocomputing, 74(16):2566–2572, 2011

Junfa Liu, Yiqiang Chen, Mingjie Liu, and Zhongtang Zhao. Selm: Semi-supervised elm with application in sparse calibrated location estimation.Neurocomputing, 74(16):2566–2572, 2011

work page 2011
[40]

Sparse bayesian extreme learning machine for multi-classification

Jiahua Luo, Chi-Man V ong, and Pak-Kin Wong. Sparse bayesian extreme learning machine for multi-classification. IEEE Transactions on Neural Networks and Learning Systems, 25(4):836–843, 2013

work page 2013
[41]

Sparse extreme learning machine for classification.IEEE transactions on cybernetics, 44(10):1858–1870, 2014

Zuo Bai, Guang-Bin Huang, Danwei Wang, Han Wang, and M Brandon Westover. Sparse extreme learning machine for classification.IEEE transactions on cybernetics, 44(10):1858–1870, 2014

work page 2014
[42]

Sparse coding extreme learning machine for classification.Neurocomputing, 261:50–56, 2017

Yuanlong Yu and Zhenzhen Sun. Sparse coding extreme learning machine for classification.Neurocomputing, 261:50–56, 2017

work page 2017
[43]

JHU press, 2013

Gene H Golub and Charles F Van Loan.Matrix computations. JHU press, 2013

work page 2013
[44]

Low-rank matrix approximation using the lanczos bidiagonalization process with applications.SIAM Journal on Scientific Computing, 21(6):2257–2274, 2000

Horst D Simon and Hongyuan Zha. Low-rank matrix approximation using the lanczos bidiagonalization process with applications.SIAM Journal on Scientific Computing, 21(6):2257–2274, 2000

work page 2000
[45]

Lsqr: An algorithm for sparse linear equations and sparse least squares.ACM Transactions on Mathematical Software (TOMS), 8(1):43–71, 1982

Christopher C Paige and Michael A Saunders. Lsqr: An algorithm for sparse linear equations and sparse least squares.ACM Transactions on Mathematical Software (TOMS), 8(1):43–71, 1982

work page 1982
[46]

Lsmr: An iterative algorithm for sparse least-squares problems

David Chin-Lung Fong and Michael Saunders. Lsmr: An iterative algorithm for sparse least-squares problems. SIAM Journal on Scientific Computing, 33(5):2950–2971, 2011. University of British Columbia

work page 2011