Pith. sign in

REVIEW 3 major objections 7 minor 43 references

Muti-Fidelity Prediction and Uncertainty Quantification with Laplace Neural Operators for Parametric Partial Differential Equations

T0 review · 3 major / 7 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read MF-LNO trains a Laplace neural operator on low-fidelity data, adds parallel linear and nonlinear high-fidelity correctors with a learnable weight, and reports 40–80 percent lower test losses than single-fidelity baselines.

desk verdict Solid incremental MF operator learning paper; the big caveat is that every LF-HF relation is synthetic and globally affine, so the impressive loss reductions don't yet transfer to realistic fidelity gaps. read the letter →

arxiv 2502.00550 v1 pith:2QKSUBC7 submitted 2025-02-01 cs.LG cs.NAmath.NAphysics.comp-ph

Haoyang Zheng , Guang Lin This is my paper · ORCID
classification cs.LGcs.NAmath.NAphysics.comp-ph
keywords Laplaceneuraloperatorsmulti-fidelitymodelingoperatorlearninguncertaintyquantificationreplicaexchangestochasticgradientLangevindynamicsparametricPDEssurrogate
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

This paper is trying to show that a neural operator can learn accurate high-fidelity surrogates for parametric differential equations when high-fidelity training data are extremely scarce, by leaning on abundant low-fidelity data. It proposes MF-LNO: a low-fidelity LNO base model, two parallel LNO correctors (linear and nonlinear), a trainable weight between them, and replica-exchange stochastic gradient Langevin dynamics for uncertainty quantification. If the claim holds, expensive high-fidelity simulations no longer need to be the main training signal; a cheap low-fidelity solver plus a handful of high-fidelity runs can build a surrogate with calibrated error bars. The paper reports test-loss reductions of 40–80% over single-fidelity baselines on four benchmark systems.

What carries the argument

The workhorse is the Laplace layer, which represents kernels and lifted inputs in pole-residue form in the Laplace domain so that convolution becomes a sum over residues at trainable poles; this supplies the frequency-domain structure of each LNO. The multi-fidelity mechanism stacks three LNOs: $\mathcal{G}_L$ maps inputs to low-fidelity outputs, then the concatenated pair $(f, \mathcal{G}_L(f))$ feeds a linear corrector $\mathcal{G}_l$ and a nonlinear corrector $\mathcal{G}_{nl}$, combined as $\hat u_H = \alpha \mathcal{G}_l(f,\mathcal{G}_L(f)) + (1-\alpha)\mathcal{G}_{nl}(f,\mathcal{G}_L(f))$ with learnable $\alpha$. Training proceeds in two phases, freezing the low-fidelity network before fitting the correctors, and uncertainty quantification comes from reSGLD, which runs multiple chains at different temperatures with deterministic even-odd swaps and adaptive correction buffers.

What would settle it

Generate low-fidelity data by coarsening a fine-mesh numerical solver (rather than adding a hand-crafted bias like $u_L = u_H + at + b$), keep only 10 high-fidelity training samples, and check whether MF-LNO's test loss still beats an LNO trained only on high-fidelity data; if the localized discretization error overwhelms the correctors, the central claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that a multi-fidelity Laplace neural operator (MF-LNO) can approximate the operator mapping input functions to high-fidelity solutions of parametric ODEs and PDEs when high-fidelity data are scarce, by first learning the low-fidelity map and then correcting it with two parallel LNOs—one without activation (linear) and one with (nonlinear)—whose weighted sum, with a trainable $\alpha$, captures the additive or multiplicative gap between fidelities. Optimized with replica-exchange SGLD, the model produces predictive means and 95% confidence intervals. Across the Lorenz system, Duffing oscillator, Burgers equation, and Brusselator reaction-diffusion system, the paper reports testing losses reduced by roughly 40–80% compared with LNOs trained on low-fidelity, high-fidelity, or mixed data alone.

Load-bearing premise

The load-bearing premise is that the gap between low- and high-fidelity data is simple enough to be captured by fixed linear or nonlinear corrections learned from a handful of high-fidelity samples.

Editorial extensions

If this is right

  • With as few as 10 high-fidelity training samples (and 1 for Brusselator), MF-LNO achieves lower test loss than LNOs trained on high-fidelity data alone, low-fidelity data alone, or a simple mix of both.
  • The trainable weight $\alpha$ lets the model interpolate between linear and nonlinear fidelity corrections, which the paper shows matters when low-fidelity data are produced by additive versus multiplicative distortions.
  • The reSGLD ensemble provides both a predictive mean and a 95% confidence interval for every test case, so the surrogate can flag regions where the data do not support a confident prediction.
  • The two-phase training schedule—train the low-fidelity network first, then freeze it and train the correctors—prevents the scarce high-fidelity data from destroying the low-fidelity representation.

Reading between the lines

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

  • Because all four experiments create low-fidelity data from high-fidelity data by global formulas like $u_L = u_H + at + b$, the paper does not yet test localized or parameter-dependent fidelity gaps; a natural next benchmark is a coarse-mesh/fine-mesh pair where the discrepancy concentrates at shocks or boundary layers.
  • The reported 40–80% reductions compare against single-fidelity LNO baselines; whether the same margin holds against other multi-fidelity methods, such as autoregressive Gaussian processes, is not established here.
  • The 95% intervals are shown visually but not calibrated; a holdout coverage test would tell whether the uncertainty estimates are reliable enough for decision-making.
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

3 major / 7 minor

Summary. This paper proposes MF-LNO, a multi-fidelity extension of Laplace Neural Operators. A low-fidelity LNO is trained first, then two parallel LNO correctors (one linear, one nonlinear) with a learnable weight alpha refine the HF prediction; training uses a replica-exchange SGLD sampler (reSGLD). The method is evaluated on four synthetic benchmarks (Lorenz, Duffing, Burgers, Brusselator) with LF data generated from HF data by simple additive or multiplicative formulas. The paper reports 40-80% reductions in testing loss relative to single-fidelity LNO baselines, with only 10 (or 1) HF training samples, and displays 95% confidence intervals from ensemble sampling. Ablation studies examine learning rate, number of chains, and network widths.

Significance. If the empirical claims hold, MF-LNO would be a simple and useful recipe for multi-fidelity operator learning with sparse HF data. The paper is clearly written, includes 10-repeat statistics, and provides ablation studies that give practical guidance on hyperparameter selection. The LNO background material in Section 2.1 is helpful. However, the significance is limited by the synthetic construction of the LF-HF relationship, which matches the corrector form by design, and by the absence of any calibration check for the reported confidence intervals. The comparison set is restricted to LNO variants, so the generality of the 40-80% claim is untested against other multi-fidelity surrogates.

major comments (3)
  1. [Section 2.2, Section 3.2, Table 2] The central empirical claim of 40-80% testing-loss reductions rests on four experiments in which the low-fidelity data are generated from the high-fidelity data by global, smooth formulas that coincide exactly with the functional forms the linear and nonlinear correctors can represent: uL = uH + at + b for Lorenz and Duffing, uL = uH + a1t + a2x + b for Burgers, uL = uH + a1t + a2x + a3y + b for Brusselator, and uL = auH sin(2πbt) for the Lorenz nonlinear case. The paper's own assumption in Section 2.2 — that the inter-fidelity relationship is 'sufficiently simple to be modeled with limited data' — is therefore satisfied by construction. The experiments do not test whether the method degrades gracefully when the LF-HF discrepancy is spatially localized, changes non-smoothly with the input parameter, or arises from a genuinely different discretization or physical approximation. I recommend adding at least one experiment with LF data generated by a coarser solver for the same PDE (or a reduced model) and reporting the relative L2 losses for LF, HF, Mix, and MF predictions; this would directly probe the validity of the assumption in Section 2.2.
  2. [Section 2.3, Section 3.2 (Brusselator), Figs. 3-7] The uncertainty quantification claim is not validated. The paper displays 95% confidence intervals in Figures 3-7 but never reports empirical coverage or interval sharpness on held-out test samples, so the reader cannot assess whether the intervals are calibrated. In addition, the Brusselator experiment's Phase 2 uses a 1-chain Adam SGLD optimizer ('we employ a 1-chain Adam SGLD optimizer for 500 epochs'), so the replica-exchange mechanism that is claimed to provide the posterior distribution is not actually used in that experiment. Please add a calibration study (e.g., empirical coverage of the 95% intervals across the test set) and either run the Brusselator Phase 2 with multiple reSGLD chains or explicitly state that the UQ for that benchmark is not produced by reSGLD.
  3. [Section 3.1, Table 2] The comparison set contains only LNO variants (LF, HF, and Mix predictions). The abstract and Discussion claim 'significant improvements... compared to traditional approaches,' but no comparison is made with established multi-fidelity surrogate methods, such as multi-fidelity DeepONet (Ref. [16]) or multi-fidelity Gaussian process regression (Ref. [17]). Without such a comparison, the reader cannot tell whether the reported gains are due to the specific MF-LNO architecture or to the general benefit of multi-fidelity training. I suggest adding at least one non-LNO multi-fidelity baseline, or rephrasing the claim to refer specifically to single-fidelity LNO baselines.
minor comments (7)
  1. [Title] The title misspells 'Multi-Fidelity' as 'Muti-Fidelity'; please correct.
  2. [Section 3.2, Burgers paragraph] 'learn the mapping from fL(x, t) to xL(x, t)' should read 'uL(x, t)' for consistency with the notation introduced in Section 2.2.
  3. [Section 3.2, Lorenz nonlinear case] The relation 'fL(t) → auH(t) sin(2πbt)' should be written as 'uL(t) = a uH(t) sin(2πbt)' to clarify that the LF output is defined as a function of the HF output.
  4. [Section 2.1, Eq. (8)] The infinite sum over ℓ in Eq. (8) is truncated in practice, but the text does not state the truncation order; please state the number of modes used (consistent with the N modes in Eq. (4)) or replace the sum with a finite-sum expression.
  5. [Section 2.3, Eq. (14)] The moment update equations in Eq. (14) use indices t-1 on the right-hand side while the surrounding text and Eq. (15) use i; please make the index notation consistent (e.g., m_{i-1} and V_{i-1}).
  6. [Figure 11] The y-axis of Fig. 11 is labeled 'Validation Loss,' but the metric is not specified; please state that it is the relative L2 loss used elsewhere in the paper.
  7. [Table 2 vs. Figure 9 caption] The Brusselator LF prediction loss is reported as 44.842±49.013 in Table 2 but 44.361±4.901 in the Figure 9 caption; these numbers should be reconciled.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported 40-80% test-loss reductions are empirical results on held-out test sets, not derived from fitted parameters or self-cited premises.

full rationale

The paper's central claim is an empirical performance comparison: testing losses are measured on held-out HF test data after training on separate HF/LF training sets (Table 2, Section 3.2). The reported quantities are not fitted parameters renamed as predictions; the loss in Eq. (12) is computed on test outputs that did not enter training, so the improvement over baselines is not forced by construction. The multi-fidelity predictor in Eq. (10) is a learned combination of two LNO correctors with a trainable weight alpha; alpha and all network parameters are optimized against Eqs. (11)-(13), and no equation in the paper reduces the HF prediction to the LF input or to a fitted correction formula. Although the LF datasets are generated from HF data using explicit additive/multiplicative formulas (e.g., uL = uH + at + b for Lorenz and Duffing, and uL = uH + a1t + a2x + b for Burgers), those formulas are not supplied to the model as analytical corrections; the model receives only input-output samples, and the test losses are genuinely out-of-sample. The match between the generative form and the linear/nonlinear corrector is a threat to external validity or 'benchmark realism' rather than circularity, because the success is still measured rather than derived. The reSGLD sampler and its even-odd gradient-based swap mechanism are taken from prior works with overlapping authors ([23,24,37,38,41,42]), but these citations are used as background sampling machinery and are not invoked as the justification for the central accuracy claim; no uniqueness theorem or prior result is used to declare the MF-LNO choice forced. Therefore the claimed derivation chain is self-contained in the relevant sense: the main result is an empirical finding, not an input-equivalent derivation.

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

The method introduces no new physical entities. Its contributions are algorithmic. The main free parameters are the learnable mixing weight alpha and the many trainable weights in the LNO blocks; the hand-tuned reSGLD hyperparameters (temperatures, swap window, correction buffer) also affect results, as shown in ablation studies.

free parameters (4)
  • alpha = learned, initialized 0.50
    Blends linear and nonlinear HF correctors in Eq. (10); fitted to HF data during training.
  • LNO pole-residue parameters (poles mu_n, residues beta_n) = trained
    Part of the Laplace layers; fitted to data during training; many parameters.
  • Temperature ladder endpoints = 1e-5 to 1e-4 (geometric)
    Chosen by hand, following [35]; sensitivity noted in ablations.
  • Swap window W* and target swap rate S = W*=50, S=0.10
    Set by hand; W*=50 > theoretical min 30.
assumptions (4)
  • domain assumption The inter-fidelity map between LF and HF outputs is simple enough to be captured by a linear and a nonlinear LNO trained on limited HF data.
    Section 2.2: 'the inter-fidelity relationship... is sufficiently simple to be modeled with limited data'. This is the core premise of the method.
  • standard math The Laplace domain pole-residue representation (Eqs. 3-8) correctly approximates the convolution operator and invertibility holds.
    Taken from previous LNO work [9]; relies on residue theorem and inverse Laplace transform.
  • standard math SGLD with decaying learning rate converges to the posterior (Robbins-Monro conditions), and reSGLD mixing improves posterior approximation.
    Section 2.3 and Remark 2.2; the UQ validity depends on this.
  • ad hoc to paper The generated synthetic LF data is representative of real low-fidelity data.
    LF datasets constructed by authors using simple additive formulas in all four benchmarks (Section 3.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Muti-Fidelity Prediction and Uncertainty Quantification with Laplace Neural Operators for Parametric Partial Differential Equations." pith.science (2026). https://pith.science/paper/2QKSUBC7

@misc{pith2026250200550,
  author       = {Pith},
  title        = {Pith review of: Muti-Fidelity Prediction and Uncertainty Quantification with Laplace Neural Operators for Parametric Partial Differential Equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2QKSUBC7}},
  note         = {Machine review of arXiv:2502.00550}
}
read the original abstract

Laplace Neural Operators (LNOs) have recently emerged as a promising approach in scientific machine learning due to the ability to learn nonlinear maps between functional spaces. However, this framework often requires substantial amounts of high-fidelity (HF) training data, which is often prohibitively expensive to acquire. To address this, we propose multi-fidelity Laplace Neural Operators (MF-LNOs), which combine a low-fidelity (LF) base model with parallel linear/nonlinear HF correctors and dynamic inter-fidelity weighting. This allows us to exploit correlations between LF and HF datasets and achieve accurate inference of quantities of interest even with sparse HF data. We further incorporate a modified replica exchange stochastic gradient Langevin algorithm, which enables a more effective posterior distribution estimation and uncertainty quantification in model predictions. Extensive validation across four canonical dynamical systems (the Lorenz system, Duffing oscillator, Burgers equation, and Brusselator reaction-diffusion system) demonstrates the framework's effectiveness. The results show significant improvements, with testing losses reduced by 40% to 80% compared to traditional approaches. This validates MF-LNO as a versatile tool for surrogate modeling in parametric PDEs, offering significant improvements in data efficiency and uncertainty-aware prediction.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages

  1. [16]

    Multifidelity deep operator networks for data-driven and physics-informed problems

    Amanda A Howard, Mauro Perego, George Em Karniadakis, Panos Stinis. Multifidelity deep operator networks for data-driven and physics-informed problems. Journal of Computational Physics. 2023;493:112462

    work page 2023

  2. [17]

    Gaussian Processes for Regression

    Christopher Williams, Carl Rasmussen. Gaussian Processes for Regression. Advances in Neural Information Processing Systems. 1995;8

    work page 1995

  3. [1]

    The Finite Element Method: Linear Static and Dynamic Finite Element Analysis

    Thomas JR Hughes. The Finite Element Method: Linear Static and Dynamic Finite Element Analysis. Courier Corporation; 2003

    work page 2003

  4. [2]

    Finite Di fference Methods for Ordinary and Partial Di fferential Equations: Steady-State and Time-Dependent Problems

    Randall J LeVeque. Finite Di fference Methods for Ordinary and Partial Di fferential Equations: Steady-State and Time-Dependent Problems. SIAM; 2007

    work page 2007

  5. [3]

    Physics-Informed Neural Networks: A Deep Learning Framework for Solving Forward and Inverse Problems Involving Nonlinear Partial Differential Equations

    Maziar Raissi, Paris Perdikaris, George E Karniadakis. Physics-Informed Neural Networks: A Deep Learning Framework for Solving Forward and Inverse Problems Involving Nonlinear Partial Differential Equations. Journal of Computational physics. 2019;378:686–707

    work page 2019

  6. [4]

    Physics-Informed Machine Learning

    George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, Liu Yang. Physics-Informed Machine Learning. Nature Reviews Physics. 2021;3(6):422–440

    work page 2021

  7. [5]

    Learning Nonlinear Operators via Deeponet Based on the Universal Approximation Theorem of Operators

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, George Em Karniadakis. Learning Nonlinear Operators via Deeponet Based on the Universal Approximation Theorem of Operators. Nature Machine Intelligence. 2021;3(3):218–229

    work page 2021

  8. [6]

    Deepxde: A Deep Learning Library for Solving Differential Equations

    Lu Lu, Xuhui Meng, Zhiping Mao, George Em Karniadakis. Deepxde: A Deep Learning Library for Solving Differential Equations. SIAM Review. 2021;63(1):208–228

    work page 2021

Show all 43 references
  1. [7]

    Fourier Neural Operator for Parametric Partial Differential Equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhat- tacharya, Andrew Stuart, et al. Fourier Neural Operator for Parametric Partial Differential Equations. International Conference on Learning Representation. 2021

    2021

  2. [8]

    Neural Operator: Learning Maps Between Function Spaces with Applications to PDEs

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhat- tacharya, Andrew Stuart, et al. Neural Operator: Learning Maps Between Function Spaces with Applications to PDEs. Journal of Machine Learning Research. 2023;24(89):1–97

    2023

  3. [9]

    Laplace Neural Operator for Solving Differential Equations

    Qianying Cao, Somdatta Goswami, George Em Karniadakis. Laplace Neural Operator for Solving Differential Equations. Nature Machine Intelligence. 2024;6(6):631–640. 27

    2024

  4. [10]

    Nonlocal Kernel Network (NKN): A Stable and Resolution-Independent Deep Neural Network

    Huaiqian You, Yue Yu, Marta D’Elia, Tian Gao, Stewart Silling. Nonlocal Kernel Network (NKN): A Stable and Resolution-Independent Deep Neural Network. Journal of Computational Physics. 2022;469:111536

    2022

  5. [11]

    Pole-Residue Method for Numerical Dynamic Analysis

    Sau-Lon James Hu, Fushun Liu, Bin Gao, Huajun Li. Pole-Residue Method for Numerical Dynamic Analysis. Journal of Engineering Mechanics. 2016;142(8):04016045

    2016

  6. [12]

    Extraction of Mechanical Properties of Materials Through Deep Learning from Instrumented Indentation

    Lu Lu, Ming Dao, Punit Kumar, Upadrasta Ramamurty, George Em Karniadakis, Subra Suresh. Extraction of Mechanical Properties of Materials Through Deep Learning from Instrumented Indentation. Proceedings of the National Academy of Sciences. 2020;117(13):7052–7062

    2020

  7. [13]

    Multifidelity Deep Neural Operators for Efficient Learning of Partial Differential Equations with Applica- tion to Fast Inverse Design of Nanoscale Heat Transport

    Lu Lu, Raphaël Pestourie, Steven G Johnson, Giuseppe Romano. Multifidelity Deep Neural Operators for Efficient Learning of Partial Differential Equations with Applica- tion to Fast Inverse Design of Nanoscale Heat Transport. Physical Review Research. 2022;4(2):023210

    2022

  8. [14]

    Review of Multi-fidelity Models

    M Giselle Fernández-Godino. Review of Multi-fidelity Models. Advances in Computational Science and Engineering. 2023;1(4):351–400

    2023

  9. [15]

    A Composite Neural Network That Learns from Multi-fidelity Data: Application to Function Approximation and Inverse PDE Problems

    Xuhui Meng, George Em Karniadakis. A Composite Neural Network That Learns from Multi-fidelity Data: Application to Function Approximation and Inverse PDE Problems. Journal of Computational Physics. 2020;401:109020

    2020

  10. [18]

    Neural-Net-Induced Gaussian Process Regression for Function Approximation and Pde Solution

    Guofei Pang, Liu Yang, George Em Karniadakis. Neural-Net-Induced Gaussian Process Regression for Function Approximation and Pde Solution. Journal of Computational Physics. 2019;384:270–288

    2019

  11. [19]

    Multi-fidelity Bayesian Neural Networks: Algorithms and Applications

    Xuhui Meng, Hessam Babaee, George Em Karniadakis. Multi-fidelity Bayesian Neural Networks: Algorithms and Applications. Journal of Computational Physics. 2021;438:110361

    2021

  12. [20]

    MCMC Using Hamiltonian Dynamics

    Radford M Neal, et al. MCMC Using Hamiltonian Dynamics. Handbook of Markov Chain Monte Carlo. 2011;2(11):2

    2011

  13. [21]

    Stochastic Gradient Hamiltonian Monte Carlo

    Tianqi Chen, Emily B Fox, Carlos Guestrin. Stochastic Gradient Hamiltonian Monte Carlo. In: International Conference on Machine Learning; 2014

    2014

  14. [22]

    The No-U-Turn Sampler: Adaptively Setting Path Lengths in Hamiltonian Monte Carlo

    Matthew D Hoffman, Andrew Gelman, et al. The No-U-Turn Sampler: Adaptively Setting Path Lengths in Hamiltonian Monte Carlo. Journal of Machine Learning Research. 28 2014;15(1):1593–1623

    2014

  15. [23]

    Non-convex Learning via Replica Exchange Stochastic Gradient MCMC

    Wei Deng, Qi Feng, Liyao Gao, Faming Liang, Guang Lin. Non-convex Learning via Replica Exchange Stochastic Gradient MCMC. In: International Conference on Machine Learning. PMLR; 2020. p. 2474–2483

    2020

  16. [24]

    Exploring Non-Convex Discrete Energy Landscapes: A Langevin-Like Sampler with Replica Exchange

    Haoyang Zheng, Ruqi Zhang, Guang Lin. Exploring Non-Convex Discrete Energy Landscapes: A Langevin-Like Sampler with Replica Exchange. arXiv preprint arXiv:250117323. 2025

    2025

  17. [25]

    Riemann Manifold Langevin and Hamiltonian Monte Carlo Methods

    Mark Girolami, Ben Calderhead. Riemann Manifold Langevin and Hamiltonian Monte Carlo Methods. Journal of the Royal Statistical Society Series B: Statistical Methodology. 2011;73(2):123–214

    2011

  18. [26]

    Preconditioned Stochas- tic Gradient Langevin Dynamics for Deep Neural Networks

    Chunyuan Li, Changyou Chen, David Carlson, Lawrence Carin. Preconditioned Stochas- tic Gradient Langevin Dynamics for Deep Neural Networks. In: AAAI Conference on Artificial Intelligence. vol. 30; 2016

    2016

  19. [27]

    An Adaptive Empirical Bayesian Method for Sparse Deep Learning

    Wei Deng, Xiao Zhang, Faming Liang, Guang Lin. An Adaptive Empirical Bayesian Method for Sparse Deep Learning. In: Advances in Neural Information Processing Systems; 2019

    2019

  20. [28]

    Underdamped Langevin MCMC: A Non-asymptotic Analysis

    Xiang Cheng, Niladri S Chatterji, Peter L Bartlett, Michael I Jordan. Underdamped Langevin MCMC: A Non-asymptotic Analysis. In: Conference on Learning Theory. PMLR; 2018. p. 300–323

    2018

  21. [29]

    Improved Discretization Analysis for Underdamped Langevin Monte Carlo

    Shunshi Zhang, Sinho Chewi, Mufan Li, Krishna Balasubramanian, Murat A Erdogdu. Improved Discretization Analysis for Underdamped Langevin Monte Carlo. In: Conference on Learning Theory. PMLR; 2023. p. 36–71

    2023

  22. [30]

    Accelerating Approximate Thompson Sampling with Underdamped Langevin Monte Carlo

    Haoyang Zheng, Wei Deng, Christian Moya, Guang Lin. Accelerating Approximate Thompson Sampling with Underdamped Langevin Monte Carlo. In: Proceedings of the International Workshop on Artificial Intelligence and Statistics; 2024

    2024

  23. [31]

    Cyclical Stochastic Gradient MCMC for Bayesian Deep Learning

    Ruqi Zhang, Chunyuan Li, Jianyi Zhang, Changyou Chen, Andrew Gordon Wilson. Cyclical Stochastic Gradient MCMC for Bayesian Deep Learning. In: International Conference on Learning Representation; 2020

    2020

  24. [32]

    A Complete Recipe for Stochastic Gradient MCMC

    Yi-An Ma, Tianqi Chen, Emily Fox. A Complete Recipe for Stochastic Gradient MCMC. Advances in Neural Information Processing Systems. 2015;28

    2015

  25. [33]

    Log-Concave Sampling: Metropolis-Hastings Algorithms Are Fast

    Raaz Dwivedi, Yuansi Chen, Martin J Wainwright, Bin Yu. Log-Concave Sampling: Metropolis-Hastings Algorithms Are Fast. Journal of Machine Learning Research. 2019;20(183):1–42

    2019

  26. [34]

    Optimal Dimension Dependence of the Metropolis-Adjusted Langevin Algorithm

    Sinho Chewi, Chen Lu, Kwangjun Ahn, Xiang Cheng, Thibaut Le Gouic, Philippe Rigol- let. Optimal Dimension Dependence of the Metropolis-Adjusted Langevin Algorithm. 29 In: Conference on Learning Theory. PMLR; 2021. p. 1260–1300

    2021

  27. [35]

    Stochastic Gradient Langevin Dynamics with Adaptive Drifts

    Sehwan Kim, Qifan Song, Faming Liang. Stochastic Gradient Langevin Dynamics with Adaptive Drifts. Journal of Statistical Computation and Simulation. 2022;92(2):318–336

    2022

  28. [36]

    Provable and Practical: Efficient Exploration in Reinforcement Learning via Langevin Monte Carlo

    Haque Ishfaq, Qingfeng Lan, Pan Xu, A Rupam Mahmood, Doina Precup, Anima Anand- kumar, et al. Provable and Practical: Efficient Exploration in Reinforcement Learning via Langevin Monte Carlo. In: International Conference on Learning Representation; 2024

    2024

  29. [37]

    Replica Exchange for Non-Convex Optimization

    Jing Dong, Xin T Tong. Replica Exchange for Non-Convex Optimization. Journal of Machine Learning Research. 2021;22:1–59

    2021

  30. [38]

    Accelerating Convergence of Replica Exchange Stochastic Gradient Mcmc via Variance Reduction

    Wei Deng, Qi Feng, Georgios Karagiannis, Guang Lin, Faming Liang. Accelerating Convergence of Replica Exchange Stochastic Gradient Mcmc via Variance Reduction. International Conference on Learning Representation. 2020

    2020

  31. [39]

    Bayesian Learning via Stochastic Gradient Langevin Dynamics

    Max Welling, Yee W Teh. Bayesian Learning via Stochastic Gradient Langevin Dynamics. In: International Conference on Machine Learning. Citeseer; 2011. p. 681–688

    2011

  32. [40]

    Stochastic Gradient Hamiltonian Monte Carlo

    Tianqi Chen, Emily Fox, Carlos Guestrin. Stochastic Gradient Hamiltonian Monte Carlo. In: International Conference on Machine Learning. PMLR; 2014. p. 1683–1691

    2014

  33. [41]

    Non-reversible Parallel Tempering for Deep Posterior Approximation

    Wei Deng, Qian Zhang, Qi Feng, Faming Liang, Guang Lin. Non-reversible Parallel Tempering for Deep Posterior Approximation. In: AAAI Conference on Artificial Intelligence. vol. 37; 2023. p. 7332–7339

    2023

  34. [42]

    Constrained Explo- ration via Reflected Replica Exchange Stochastic Gradient Langevin Dynamics

    Haoyang Zheng, Hengrong Du, Qi Feng, Wei Deng, Guang Lin. Constrained Explo- ration via Reflected Replica Exchange Stochastic Gradient Langevin Dynamics. In: International Conference on Machine Learning. vol. 235. PMLR; 2024. p. 61321–61348

    2024

  35. [43]

    Delving Deep into Rectifiers: Surpassing Human-Level Performance on Imagenet Classification

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. Delving Deep into Rectifiers: Surpassing Human-Level Performance on Imagenet Classification. In: Conference on Computer Vision and Pattern Recognition; 2015. p. 1026–1034. 30

    2015

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.