GNet: A scalable and flexible Gaussian process network with nonparametric neurons

Mengyang Gu

REVIEW 3 major objections 5 minor 45 references

GNet: A scalable and flexible Gaussian process network with nonparametric neurons

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read GNet replaces neural-network activations with Gaussian processes and uses a new jointly inverse Kalman filter to compute covariance-vector products in O(nq^2), making GP-based networks trainable and predictable without ever forming a covari

desk verdict JIKF is a real O(n q^2) contribution and the one-layer model holds up, but the deep model's fixed 10-knot approximation and a suspicious NRMSE formula mean the multi-layer claims should be read with caution. read the letter →

arxiv 2607.10735 v2 pith:OALOF2K4 submitted 2026-07-12 stat.ME stat.COstat.ML

Mengyang Gu This is my paper

classification stat.MEstat.COstat.ML MSC 62G0862M2068T0760G15

keywords GaussianprocessnetworknonparametricneuronsjointlyinverseKalmanfilterdynamiclinearmodelsMatérnkernelclosed-formgradientsscalablepredictiondensityfunctionaltheory

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 proposes GNet, a network in which each neuron's activation function is a Gaussian process with a one-dimensional input, and shows that such networks can be trained and used at scale. The key enabler is the jointly inverse Kalman filter (JIKF), which computes the product of a DLM-induced covariance matrix with any vector in O(nq^2) operations instead of O(n^2), given commuting transition matrices. Closed-form gradients are also derived, so optimization and prediction avoid forming the n-by-n covariance matrix altogether. In experiments with up to about 95,000 training points and 201-dimensional inputs, GNet models achieve lower held-out normalized RMSE than neural networks, Kolmogorov-Arnold networks, exact Gaussian processes (at small n), and Vecchia-type approximations in most tested scenarios.

What carries the argument

The jointly inverse Kalman filter (JIKF): two recurrence relations (Eqs. 15–16) that compute the DLM covariance-vector product b = Σv in O(nq^2) time, where q is the latent-state dimension, assuming the transition matrices commute (Eq. 10). It replaces direct O(n^2) matrix-vector multiplication and, combined with closed-form gradients, lets the entire GNet training loop avoid forming the n-by-n covariance matrix. The filter's correctness rests on Lemma 1's factorization of cross-covariances into products of transition matrices times the stationary latent covariance.

What would settle it

For a dynamic linear model whose transition matrices commute with each other but not with the stationary covariance W_1, compute Σv directly and via the JIKF recursions; any mismatch would falsify Theorem 1 as stated. Alternatively, in a deep GNet, increase the knot count m_l,K from 10 to 50 and check whether held-out NRMSE changes materially.

Watch

Extended reading notes

Core claim

The paper's central claim is that a Gaussian process network with nonparametric neurons is computationally tractable at large scale when the activation kernels come from dynamic linear models. The proof mechanism is the jointly inverse Kalman filter: because the DLM transition matrices commute, the covariance-vector product Σv can be computed exactly by two O(nq^2) recursions, one forward and one backward, without assembling Σ. Together with closed-form loss gradients (Lemmas 2 and 3), training and prediction become matrix-free. Empirically, under a unified optimization setting, the resulting GNet models produce lower held-out normalized RMSE than several mainstream baselines across a range

Load-bearing premise

The recursions are exact only if the DLM transition matrices commute with each other and with the stationary latent covariance; the paper explicitly assumes the former but not the latter, and for deep GNet the hidden-layer activations are approximated by 10 knots, so the deep model's accuracy is heuristic.

Editorial extensions

If this is right

A pretrained GNet can predict each new input in time that grows only logarithmically with the training sample size, making it feasible to serve many concurrent users.
Closed-form gradients let practitioners train GNet without automatic differentiation through iterative filters, avoiding excessive memory and computational graphs.
GNet brings GP-based predictive modeling to sample sizes around 10^5 with high-dimensional inputs, a regime where exact Gaussian processes are computationally prohibitive.
Because the final layer is an exact Gaussian process, GNet provides predictive intervals at modest extra cost, unlike standard neural-network baselines in the study.
The 10-knot representation of hidden layers (Eq. 25) is a concrete approximation that can be refined or adaptively placed to improve deep GNet accuracy.

Reading between the lines

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

If the JIKF recursions were extended to non-commuting DLM transition families, the same O(nq^2) acceleration could apply to a broader class of state-space and kernel methods, but the paper does not investigate that case.
The strong empirical results for a one-layer, 10-neuron GNet on high-dimensional physical inputs suggest that one-dimensional nonparametric feature maps may be a sufficient inductive bias for certain functionals; testing GNet on other physical quantities would clarify this.
The deep GNet's 10-knot approximation is a pragmatic heuristic; varying the knot count (e.g., from 5 to 50) would provide a sensitivity analysis and reveal bias-variance tradeoffs in new applications.
The paper explicitly leaves reliable uncertainty quantification for deep layers as an open problem; future work could compare GNet's predictive interval coverage with exact Bayesian approaches on a range of datasets.

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.

Desk Editor's Note

read the letter

The JIKF algorithm is genuine, the one-layer GNet is on solid ground, and the deep model is the soft spot. The pairwise commutation condition in Theorem 1 is enough; the \tilde W_1 concern in the stress-test doesn't hold up, since \tilde W_1 always sits on the right in Lemma 1. The recursions are simple and correct for DLM covariances with commuting transition matrices, and the O(nq^2) cost plus closed-form gradients is a real contribution. The Nyström-PCG integration is sensible, and the log(n) prediction cost for pretrained models is useful.

The numeric work is broad: simulated functions, UCI regressions, and a cDFT problem with n≈95k and p=201. GNet-1L looks consistently good, and the cDFT results are impressive for a 2,031-parameter model. However, the GNet-2L results need a caveat. Section 3 replaces each hidden GP with a fixed 10-knot predictive process (Eq 25), and the predictive equations then treat the induced last-layer inputs as known. So the two-layer model is an approximation, with no error bound and no propagation of hidden-layer uncertainty. Since GNet-2L is the best method in several experiments, those numbers are evidence about the approximate model, not about an exact deep GNet. That doesn't hurt the JIKF contribution, but it should be stated explicitly and studied.

Two smaller issues. First, Eq (32)'s NRMSE denominator uses \hat y - \bar y rather than y - \bar y. That makes the scaling depend on the predictions and is non-standard; if that's what was computed, it could distort comparisons. Worth a correction. Second, no code or data are included, and some baselines use non-default hyperparameters. That's not disqualifying, but it makes the empirical claims harder to audit.

Overall, this is a technique-level contribution that deserves a serious referee. The JIKF and closed-form gradients are likely correct and practically useful. I'd engage with the paper, primarily for the one-layer model and the algorithmic machinery. The deep model needs more work before its claims can be taken at face value.

Referee Report

3 major / 5 minor

Summary. The paper introduces GNet, a Gaussian process regression network in which each neuron's activation function is a GP acting on a one-dimensional projected input. To make the model scalable, the author proposes the jointly inverse Kalman filter (JIKF), an exact O(n q^2) algorithm for DLM-induced covariance-matrix--vector products when transition matrices commute, and derives closed-form gradients for training without forming covariance matrices. Training is further accelerated by combining JIKF with a Nyström-preconditioned conjugate gradient method. The method is tested on simulated nonlinear functions, five UCI regression data sets, and a 201-dimensional classical density functional theory prediction problem, with comparisons to exact and Vecchia-approximated GPs, feedforward neural networks, and Kolmogorov--Arnold networks. The paper claims that GNet achieves lower held-out NRMSE than the non-GNet baselines in most settings and that a pretrained GNet can predict at O(log n) cost per test point.

Significance. If the claims hold, the paper makes a useful contribution. The JIKF recursions in Theorem 1 are elegant and appear correct: I checked the concern that Theorem 1 also requires each G_i to commute with \tilde W_1; this is not needed because \tilde W_1 always appears on the right-hand side of the transition products in Lemma 1, and the backward recursion (16) only requires G_{i+1} to commute with later transition matrices, which is implied by (10). The closed-form gradient derivations in Lemmas 2--3 are nontrivial and, if correct, avoid the prohibitive automatic-differentiation graphs of iterative solvers. The empirical study is broad and includes a practically motivated high-dimensional cDFT example. On the negative side, no code or data are provided, and the deep version of the model rests on an unquantified low-rank approximation that is load-bearing for several main empirical claims. The JIKF contribution itself is sufficiently solid that the concerns are addressable in a revision.

major comments (3)

[Section 3, Eq. (25), and Eqs. (5)--(7)] For L>1, the model is not a true deep GP but a heuristic finite-rank approximation. Hidden-layer GPs are replaced by 10-knot interpolants with directly optimized whitened values, and the predictive equations condition on the last-layer inputs \tilde x_L as if they were known. This ignores posterior uncertainty in the hidden layers, and no error bound or sensitivity analysis is given for m_{l,K}=10. Since GNet-2L is the best-performing variant in most of the UCI experiments (Figure 4) and in the cDFT experiment (Figure 6), the central empirical claims for the deep model depend on this unquantified approximation. Please add a knot-convergence study (e.g., m=20,50,100), state clearly that the predictive intervals cover only last-layer noise and not hidden-layer uncertainty, or reframe the deep model as an explicit approximate/parametric model with appropriate qualifications.
[Section 4, Table 1, and Eqs. (15)--(16)] The claim that a pretrained GNet predicts each test input in O(log n) time is not substantiated by the presented recursions. Equations (15)--(16) compute b=Σv for all n data locations in O(n q^2); they do not by themselves show how to evaluate a single cross-covariance sum Σ_{i} c(x*, x_i) α_i at an arbitrary test point x* in O(log n) time. The cited Section S3.1 must specify the data structure/algorithm (e.g., prefix sums with sorted inputs and binary search) and any additional assumptions (sorted inputs, commutativity, low-rank preconditioner). This is a central scalability claim for the 'many concurrent users' scenario and should be made precise.
[Eq. (32)] The definition of NRMSE is nonstandard and likely a typo: the denominator is written as Σ_i (ŷ(x_i) − ȳ)^2, using the model predictions ŷ, rather than the observed values y_i or the latent y_signal(x_i). As written, a model that predicts the constant mean has a zero denominator, and the reported normalized errors are not directly comparable to standard NRMSE in the literature. If this is a typographical error, please correct the denominator to use observed values; if intentional, justify the choice and explain why it does not distort the cross-method comparisons.

minor comments (5)

[Section 5, paragraph before Figure 2] The phrase 'For compariosn' should be 'For comparison'.
[Section 5.1, end of first paragraph] The text says 'we report the the average performance metrics'; remove the duplicated 'the'.
[Section 4, Table 1] The symbols k_pcg and m_pcg appear in the table but are not defined in the main text; define them before the table and state how they are set in the experiments.
[Section 5, benchmark setting] The GNet models use a fixed optimization setting, but some baselines are hand-tuned (e.g., FNN-R learning rates, Vecchia conditioning sizes). This is acceptable, but the text should explicitly say that the comparison is not fully automated/unified across methods and that baseline tuning was chosen to improve baseline performance, not to disadvantage them.
[General] No code or data availability statement is provided. Given the algorithmic nature of the contribution and the many empirical comparisons, releasing code would substantially strengthen reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the JIKF recursion is exact algebra from Lemma 1, and the empirical claims rest on held-out data.

full rationale

The paper's central derivation is self-contained. Theorem 1 (Eqs. 15-16) is a direct algebraic factorization of the sums in Lemma 1 (Eqs. 12-14): each backward/forward recursion factors a common transition matrix out of a sum of covariance-vector products, so the claimed O(n q^2) computation of b = Σv is an exact rearrangement, not an assumption of the result. The gradient lemmas are derivatives of the validation MSE (Eq. 17), and the reported NRMSE values are evaluated on held-out test inputs (Figures 2, 4, and 6), so no fitted parameter is renamed as a prediction. The self-citation to the inverse Kalman filter [9] is used only as a computational comparison ('smaller than a recent approach [9]'), not as a load-bearing justification. The genuine limitations noted in the paper—the fixed 10-knot approximation for hidden layers (Eq. 25), conditioning on the last-layer inputs as known (Eq. 5), and the resulting unquantified deep-model approximation error—are heuristic approximation issues rather than circular definitions or fitted-input predictions. They affect the strength of the empirical claims but do not make the derivation chain circular. Hence the appropriate score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The GNet's speed and accuracy claims are built on the DLM/Matern kernel class, the unstated W1-commutativity condition, the fixed-knot hidden-layer approximation, and the conditioning approximation for deep layers. These are not fitted to the target results but are modeling/algorithmic assumptions the reader must accept. Standard weights/scales of the model are the objects of training and are not listed as free parameters.

free parameters (4)

Matern smoothness ν=5/2 = 5/2 (fixed)
Fixed across all experiments; not tuned per dataset, but the main theory applies to any half-integer ν, so the choice is a default rather than a fitted constant.
Hidden-layer knot count m_{l,K}=10 = 10
Used for all L>1 models; larger values would improve the approximation but increase cost. The paper does not analyze sensitivity.
Nyström preconditioner size / PCG tolerance = not reported
Affects the accuracy of α=Σ^{-1}y and hence all reported predictions; the main text reports only that RMSE vs direct computation is <1e-8, not the subset size.
Adam learning rate 0.02 / min epochs 100 = 0.02 / 100
Hand-chosen optimization settings; the paper deliberately uses a unified setting rather than per-dataset tuning.

assumptions (5)

domain assumption Covariance kernel c_{l,j} is induced by a continuous-time DLM with transition matrices satisfying G_i G_{i'} = G_{i'} G_i (Eq. 10).
Restricts the GP network to the DLM/Matern half-integer kernel class; the fast algorithm does not apply to general kernels.
domain assumption The stationary covariance \tilde W_1 commutes with every G_i so that the recursions in Theorem 1 hold.
Needed to pass from Lemma 1 to the O(n q^2) recursions; not stated in the paper. Holds for Matern kernels if \tilde W_1 is a function of the same generator J, but the general claim is unsupported.
ad hoc to paper For L>1, replacing each hidden GP by 10 knots and optimized whitened values (Eq. 25) gives a sufficiently accurate approximation of the latent processes.
No approximation-error bound is given; this is the mechanism that makes deep GNet tractable.
domain assumption Predictive equations (5)-(7) condition on the last-layer transformed inputs \tilde x_L as known, ignoring uncertainty from previous layers.
For L=1 this is exact; for L>1 it is an approximation not quantified in the paper.
domain assumption Benchmark implementations (RobustGaSP, Vecchia, SVecchia, FNN, KAN) are correctly configured and their hyperparameters are comparable.
The paper sets GNet's optimization scheme uniformly but adjusts FNN learning rates and Vecchia conditioning sizes, which complicates the comparison.

how reviews work

0 comments

Cite this review

Pith. "Pith review of GNet: A scalable and flexible Gaussian process network with nonparametric neurons." pith.science (2026). https://pith.science/paper/OALOF2K4

@misc{pith2026260710735,
  author       = {Pith},
  title        = {Pith review of: GNet: A scalable and flexible Gaussian process network with nonparametric neurons},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OALOF2K4}},
  note         = {Machine review of arXiv:2607.10735}
}

read the original abstract

We develop GNet, a scalable and flexible Gaussian process network with nonparametric activation functions modeled by Gaussian processes. To reduce computational and storage costs, we introduce the jointly inverse Kalman filter, a fast algorithm together with closed-form expressions of gradients for accelerating model training and predictions without the need to form covariance matrices. Using a unified optimization setting, GNet shows competitive performance across a diverse range of test problems, including predicting nonlinear functions, nonparametric regression of real-world data, and predicting one-body direct correlation functions with high-dimensional inputs in classical density function theory. The strong performance of GNet, accelerated by the jointly inverse Kalman filter, suggests broad applicability to large-scale predictive modeling with substantially reduced computational and storage costs.

Figures

Figures reproduced from arXiv: 2607.10735 by the authors.

**Figure 1.** (a) Computational time for b “ Σ˜ Ly by direct matrix-vector multiplication and by the JIKF for the exponential kernel and Mat´ern kernel in Equations (3) and (2), respectively. Both methods are exact. The solid squares and diamonds overlap. (b) Time to compute pΣ˚ L q T Σ˜ ´1 L y for predicting n ˚ “ 104 inputs by direct computation, conjugate gradient (CG) approach with direct matrix-vector multiplication, and JIK… view at source ↗

**Figure 2.** Comparison of the NRMSE (first row) and computational time (second row) at different [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗

**Figure 2.** Comparison of the NRMSE (first row) and computational time (second row) at different [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

Figures from the paper (7 more)

**Figure 3.** Figure 3: The residual of the predictions on regular grids by different approaches in one experiment [PITH_FULL_IMAGE:figures/full_fig_p013_3.png]

**Figure 4.** Figure 4: The average NRMSE for five response variables from four UCI data sets. The total [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]

**Figure 4.** Figure 4: The average NRMSE for five response variables from four UCI data sets. The total [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]

**Figure 5.** Figure 5: (a) One-dimensional hard-rod system with particle length ˜a [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

**Figure 5.** Figure 5: (a) One-dimensional hard-rod system with particle length ˜a [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

**Figure 6.** Figure 6: (a) NRMSE by different approaches across training sample sizes. (b) Computational [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 4 linked inside Pith

[1]

Learning activation functions to improve deep neural networks.arXiv preprint arXiv:1412.6830, 2014

Forest Agostinelli, Matthew Hoffman, Peter Sadowski, and Pierre Baldi. Learning activation functions to improve deep neural networks.arXiv preprint arXiv:1412.6830, 2014

arXiv 2014
[2]

CRC Press, 2014

Sudipto Banerjee, Bradley P Carlin, and Alan E Gelfand.Hierarchical modeling and analysis for spatial data. CRC Press, 2014

2014
[3]

Automatic differentiation in machine learning: a survey.Journal of Machine Learning Research, 18(153):1–43, 2018

Atilim Gunes Baydin, Barak A Pearlmutter, Alexey Andreyevich Radul, and Jeffrey Mark Siskind. Automatic differentiation in machine learning: a survey.Journal of Machine Learning Research, 18(153):1–43, 2018

2018
[4]

Pope, and Michael Marcolini

Thomas Brooks, D. Pope, and Michael Marcolini. Airfoil Self-Noise. UCI Machine Learning Repository, 1989. DOI: https://doi.org/10.24432/C5VW2C

doi:10.24432/c5vw2c 1989
[5]

Deep Gaussian processes

Andreas Damianou and Neil D Lawrence. Deep Gaussian processes. InArtificial intelligence and statistics, pages 207–215. PMLR, 2013

2013
[6]

Hierarchical nearest- neighbor Gaussian process models for large geostatistical datasets.Journal of the American Statistical Association, 111(514):800–812, 2016

Abhirup Datta, Sudipto Banerjee, Andrew O Finley, and Alan E Gelfand. Hierarchical nearest- neighbor Gaussian process models for large geostatistical datasets.Journal of the American Statistical Association, 111(514):800–812, 2016

2016
[7]

Generalized Latin hypercube design for computer ex- periments.Technometrics, 52(4), 2010

Holger Dette and Andrey Pepelyshev. Generalized Latin hypercube design for computer ex- periments.Technometrics, 52(4), 2010

2010
[8]

OUP Oxford, 2012

James Durbin and Siem Jan Koopman.Time series analysis by state space methods, volume 38. OUP Oxford, 2012

2012

Show all 45 references

[9]

The inverse Kalman filter.Biometrika, 112(4):asaf054, 2025

Xinyi Fang and Mengyang Gu. The inverse Kalman filter.Biometrika, 112(4):asaf054, 2025

2025
[10]

Reliable emulation of complex functionals by active learning with error control.The Journal of Chemical Physics, 157(21), 2022

Xinyi Fang, Mengyang Gu, and Jianzhong Wu. Reliable emulation of complex functionals by active learning with error control.The Journal of Chemical Physics, 157(21), 2022

2022
[11]

Chapman and Hall/CRC, 2020

Robert B Gramacy.Surrogates: Gaussian process modeling, design, and optimization for the applied sciences. Chapman and Hall/CRC, 2020

2020
[12]

Mengyang Gu, Jesus Palomo, and James O. Berger. RobustGaSP: Robust Gaussian stochastic process emulation in R.The R Journal, 11(1):112–136, 2019

2019
[13]

Robust Gaussian stochastic process emulation.Annals of Statistics, 46(6A):3038–3066, 2018

Mengyang Gu, Xiaojing Wang, and James O Berger. Robust Gaussian stochastic process emulation.Annals of Statistics, 46(6A):3038–3066, 2018

2018
[14]

R package version 0.5.0

Joseph Guinness, Matthias Katzfuss, and Youssef Fahmy.GpGp: Fast Gaussian Process Computation Using Vecchia’s Approximation Functions, 2024. R package version 0.5.0

2024
[15]

A Bayesian analysis of kriging.Technometrics, 35(4):403–410, 1993

Mark S Handcock and Michael L Stein. A Bayesian analysis of kriging.Technometrics, 35(4):403–410, 1993. 17

1993
[16]

Generalized additive models.Statistical Science, 1(3):297–310, 1986

Trevor Hastie and Robert Tibshirani. Generalized additive models.Statistical Science, 1(3):297–310, 1986

1986
[17]

R package version 1.5.1

Tomasz Kalinowski, Daniel Falbel, JJ Allaire, Fran¸ cois Chollet, Yuan Tang, Wouter Van Der Bijl, Martin Studer, and Sigrid Keydana.keras3: R Interface to ’Keras’, 2026. R package version 1.5.1

2026
[18]

A general framework for Vecchia approximations of Gaussian processes.Statistical Science, 36(1):124–141, 2021

Matthias Katzfuss and Joseph Guinness. A general framework for Vecchia approximations of Gaussian processes.Statistical Science, 36(1):124–141, 2021

2021
[19]

Scaled Vecchia approximation for fast computer-model emulation.SIAM/ASA Journal on Uncertainty Quantification, 10(2):537– 554, 2022

Matthias Katzfuss, Joseph Guinness, and Earl Lawrence. Scaled Vecchia approximation for fast computer-model emulation.SIAM/ASA Journal on Uncertainty Quantification, 10(2):537– 554, 2022

2022
[20]

Covariance tapering for likelihood-based estimation in large spatial data sets.Journal of the American Statistical Association, 103(484):1545–1555, 2008

Cari G Kaufman, Mark J Schervish, and Douglas W Nychka. Covariance tapering for likelihood-based estimation in large spatial data sets.Journal of the American Statistical Association, 103(484):1545–1555, 2008

2008
[21]

The UCI machine learning reposi- tory.https://archive.ics.uci.edu, 2023

Markelle Kelly, Rachel Longjohn, and Kolby Nottingham. The UCI machine learning reposi- tory.https://archive.ics.uci.edu, 2023

2023
[22]

Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

2014 arXiv
[23]

An explicit link between Gaussian fields and Gaussian markov random fields: the stochastic partial differential equation approach

Finn Lindgren, H ˚ avard Rue, and Johan Lindstr¨ om. An explicit link between Gaussian fields and Gaussian markov random fields: the stochastic partial differential equation approach. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 73(4):423–498, 2011

2011
[24]

KAN: Kolmogorov–Arnold networks

Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljacic, Thomas Hou, and Max Tegmark. KAN: Kolmogorov–Arnold networks. InInternational con- ference on learning representations, volume 2025, pages 70367–70413, 2025

2025
[25]

Neural network with optimal neuron activation func- tions based on additive Gaussian process regression.The Journal of Physical Chemistry A, 127(37):7823–7835, 2023

Sergei Manzhos and Manabu Ihara. Neural network with optimal neuron activation func- tions based on additive Gaussian process regression.The Journal of Physical Chemistry A, 127(37):7823–7835, 2023

2023
[26]

Bayesian design and analysis of computer experiments: use of derivatives in surface prediction.Technometrics, 35(3):243–255, 1993

Max D Morris, Toby J Mitchell, and Donald Ylvisaker. Bayesian design and analysis of computer experiments: use of derivatives in surface prediction.Technometrics, 35(3):243–255, 1993

1993
[27]

PhD thesis, University of Illinois at Urbana-Champaign, 1991

Jeong Soo Park.Tuning complex computer codes to data and optimal designs. PhD thesis, University of Illinois at Urbana-Champaign, 1991

1991
[28]

Springer, 2009

Giovanni Petris, Sonia Petrone, and Patrizia Campagnoli.Dynamic linear models. Springer, 2009

2009
[29]

MIT Press, 2006

Carl Edward Rasmussen.Gaussian processes for machine learning. MIT Press, 2006

2006
[30]

DiceKriging, DiceOptim: Two R packages for the analysis of computer experiments by Kriging-based metamodeling and opti- mization.Journal of Statistical Software, 51(1):1–55, 2012

Olivier Roustant, David Ginsbourger, and Yves Deville. DiceKriging, DiceOptim: Two R packages for the analysis of computer experiments by Kriging-based metamodeling and opti- mization.Journal of Statistical Software, 51(1):1–55, 2012. 18

2012
[31]

F ALKON: An optimal large scale kernel method.Advances in neural information processing systems, 30, 2017

Alessandro Rudi, Luigi Carratino, and Lorenzo Rosasco. F ALKON: An optimal large scale kernel method.Advances in neural information processing systems, 30, 2017

2017
[32]

All emulators are wrong, many are useful, and some are more useful than others: A reproducible comparison of computer model surrogates.arXiv preprint arXiv:2512.09060, 2025

Kellin N Rumsey, Graham C Gibson, Devin Francom, and Reid Morris. All emulators are wrong, many are useful, and some are more useful than others: A reproducible comparison of computer model surrogates.arXiv preprint arXiv:2512.09060, 2025

2025 arXiv
[33]

Neural functional theory for inhomogeneous fluids: Fundamentals and applications.Proceedings of the National Academy of Sciences, 120(50):e2312484120, 2023

Florian Samm¨ uller, Sophie Hermann, Daniel de Las Heras, and Matthias Schmidt. Neural functional theory for inhomogeneous fluids: Fundamentals and applications.Proceedings of the National Academy of Sciences, 120(50):e2312484120, 2023

2023
[34]

Vecchia-approximated deep Gaus- sian processes for computer experiments.Journal of Computational and Graphical Statistics, 32(3):824–837, 2023

Annie Sauer, Andrew Cooper, and Robert B Gramacy. Vecchia-approximated deep Gaus- sian processes for computer experiments.Journal of Computational and Graphical Statistics, 32(3):824–837, 2023

2023
[35]

Deep learning in neural networks: An overview.Neural networks, 61:85– 117, 2015

J¨ urgen Schmidhuber. Deep learning in neural networks: An overview.Neural networks, 61:85– 117, 2015

2015
[36]

Accurate telemoni- toring of Parkinson’s disease progression by non-invasive speech tests.IEEE Transactions on Bio-medical Engineering, 57(4):884–893, 2009

Athanasios Tsanas, Max Little, Patrick McSharry, and Lorraine Ramig. Accurate telemoni- toring of Parkinson’s disease progression by non-invasive speech tests.IEEE Transactions on Bio-medical Engineering, 57(4):884–893, 2009

2009
[37]

Accurate quantitative estimation of energy perfor- mance of residential buildings using statistical machine learning tools.Energy and buildings, 49:560–567, 2012

Athanasios Tsanas and Angeliki Xifara. Accurate quantitative estimation of energy perfor- mance of residential buildings using statistical machine learning tools.Energy and buildings, 49:560–567, 2012

2012
[38]

Gaussian process neurons learn stochastic activation functions.arXiv preprint arXiv:1711.11059, 2017

Sebastian Urban, Marcus Basalla, and Patrick van der Smagt. Gaussian process neurons learn stochastic activation functions.arXiv preprint arXiv:1711.11059, 2017

2017 arXiv
[39]

Attention is all you need.Advances in neural information processing systems, 30, 2017

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

2017
[40]

Estimation and model identification for continuous spatial processes.Journal of the Royal Statistical Society Series B: Statistical Methodology, 50(2):297–312, 1988

Aldo V Vecchia. Estimation and model identification for continuous spatial processes.Journal of the Royal Statistical Society Series B: Statistical Methodology, 50(2):297–312, 1988

1988
[41]

West and P

M. West and P. J. Harrison.Bayesian Forecasting & Dynamic Models. Springer Verlag, 2nd edition, 1997

1997
[42]

Using the Nystr¨ om method to speed up kernel machines.Advances in neural information processing systems, 13, 2000

Christopher Williams and Matthias Seeger. Using the Nystr¨ om method to speed up kernel machines.Advances in neural information processing systems, 13, 2000

2000
[43]

Minimax-optimal nonparametric regression in high dimensions

Yun Yang and Surya T Tokdar. Minimax-optimal nonparametric regression in high dimensions. The Annals of Statistics, 43(2):652 – 674, 2015

2015
[44]

Modeling of strength of high-performance concrete using artificial neural networks

I-C Yeh. Modeling of strength of high-performance concrete using artificial neural networks. Cement and Concrete research, 28(12):1797–1808, 1998

1998
[45]

Radial neighbours for provably accurate scalable approximations of Gaussian processes.Biometrika, 111(4):1151–1167, 2024

Yichen Zhu, Michele Peruzzi, Cheng Li, and David B Dunson. Radial neighbours for provably accurate scalable approximations of Gaussian processes.Biometrika, 111(4):1151–1167, 2024. 19

2024

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

Learning activation functions to improve deep neural networks.arXiv preprint arXiv:1412.6830, 2014

Forest Agostinelli, Matthew Hoffman, Peter Sadowski, and Pierre Baldi. Learning activation functions to improve deep neural networks.arXiv preprint arXiv:1412.6830, 2014

arXiv 2014

[2] [2]

CRC Press, 2014

Sudipto Banerjee, Bradley P Carlin, and Alan E Gelfand.Hierarchical modeling and analysis for spatial data. CRC Press, 2014

2014

[3] [3]

Automatic differentiation in machine learning: a survey.Journal of Machine Learning Research, 18(153):1–43, 2018

Atilim Gunes Baydin, Barak A Pearlmutter, Alexey Andreyevich Radul, and Jeffrey Mark Siskind. Automatic differentiation in machine learning: a survey.Journal of Machine Learning Research, 18(153):1–43, 2018

2018

[4] [4]

Pope, and Michael Marcolini

Thomas Brooks, D. Pope, and Michael Marcolini. Airfoil Self-Noise. UCI Machine Learning Repository, 1989. DOI: https://doi.org/10.24432/C5VW2C

doi:10.24432/c5vw2c 1989

[5] [5]

Deep Gaussian processes

Andreas Damianou and Neil D Lawrence. Deep Gaussian processes. InArtificial intelligence and statistics, pages 207–215. PMLR, 2013

2013

[6] [6]

Hierarchical nearest- neighbor Gaussian process models for large geostatistical datasets.Journal of the American Statistical Association, 111(514):800–812, 2016

Abhirup Datta, Sudipto Banerjee, Andrew O Finley, and Alan E Gelfand. Hierarchical nearest- neighbor Gaussian process models for large geostatistical datasets.Journal of the American Statistical Association, 111(514):800–812, 2016

2016

[7] [7]

Generalized Latin hypercube design for computer ex- periments.Technometrics, 52(4), 2010

Holger Dette and Andrey Pepelyshev. Generalized Latin hypercube design for computer ex- periments.Technometrics, 52(4), 2010

2010

[8] [8]

OUP Oxford, 2012

James Durbin and Siem Jan Koopman.Time series analysis by state space methods, volume 38. OUP Oxford, 2012

2012