Pith. sign in

REVIEW 3 major objections 5 minor 48 references

A Generalized Unified Skew-Normal Process with Neural Bayes Inference

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A generalized unified skew-normal spatial process can be built by adding an independent truncated Gaussian process to a Gaussian process; the paper shows this keeps long-range correlations vanishing and supports closed-form kriging with…

desk verdict The GSUN construction is well-motivated and the neural Bayes estimator is competently built, but the paper does not establish that GSUN is a valid spatial process, and that failure is load-bearing. read the letter →

arxiv 2411.17400 v2 pith:I7C3JVI7 submitted 2024-11-26 stat.ML cs.LG

Kesen Wang , Marc G. Genton This is my paper · ORCID
classification stat.MLcs.LG
keywords unifiedskew-normaldistributionnon-GaussianspatialprocessneuralBayesestimatorgraphattentionnetworkencodertransformerkrigingtruncatedGaussianMatérncovariance
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 proposes a new class of non-Gaussian spatial fields, the generalized unified skew-normal (GSUN) process, and an amortized inference scheme for it. The central claim is that $Z(s)=W(s)+h(s)W^{+}(s)$, with $W$ an ordinary Gaussian process and $W^{+}$ an independent truncated (non-negative) Gaussian process, is a valid spatial process whose correlations vanish at large distances, avoiding a failure mode of earlier skew-normal fields. The skewness matrix $H$ in the equivalent SUN representation is diagonal and built from a common effect plus eigenvalue-weighted principal components, which makes the direction and amount of skewness interpretable. Because SUN conditional distributions are closed, kriging can be done in closed form, and a graph-attention plus transformer neural Bayes estimator provides fast parameter inference once trained. The paper also reports that GSUN is empirically distinct from Gaussian and Tukey g-and-h fields through probability integral transform diagnostics, and gives a better fit to Pb-contaminated soil data than the SUGLG process.

What carries the argument

The machinery is the convolution representation of the SUN distribution, $Y=\xi+HU+W$, where $U=(W_0\mid W_0+\tau>0)$ is a truncated multivariate normal and $W$ is an independent normal; this replaces the original SUN parameterization with an explicit skewness-direction matrix $H$. The GSUN process instantiates this with $W$ a Matérn Gaussian process, $U$ the latent truncated process with its own Matérn correlation matrix $C(\theta_2)$, and $H=\operatorname{diag}(h(s))$ with $h(s)=\delta_1+\delta_2$ times an eigenvalue-weighted average of principal components of $\Sigma(\theta_1)$ and $C(\theta_2)$. This representation is what makes conditional distributions remain SUN, hence kriging is closed-form, and it makes the skewness direction interpretable. The inference machinery is the neural Bayes estimator: a network trained by minimizing Monte Carlo Bayes risk that maps observed spatial data to parameter estimates, using GAT layers to aggregate neighbor information and an encoder transformer to capture long-range dependence.

What would settle it

Simulate the model at $n$ locations with correlated Matérn truncation and compute $\operatorname{cov}\{Z(s_i),Z(s_j)\}$ for pairs far apart; if the covariance does not decay to zero as distance grows, the vanishing-correlation claim fails. Also check whether the marginal distribution at a fixed set of locations changes when extra distant locations are added to the conditioning set; if it does, GSUN is not a well-defined spatial process.

Watch

Extended reading notes

Core claim

The central claim is that $Z(s)=W(s)+h(s)W^{+}(s)$, with $W$ a Gaussian process and $W^{+}$ a truncated (non-negative) Gaussian process independent of $W$, both with Matérn covariance, defines a valid generalized unified skew-normal spatial process. The paper argues that at large distances the covariance vanishes because both the Matérn covariance of $W$ and the covariance of the truncated field $W^{+}$ vanish, so the process avoids the non-ergodicity that plagues earlier skew-normal spatial fields. It further claims that the re-parameterization $Y=\xi+HU+W$ makes the skewness matrix $H$ a diagonal matrix of location-specific weights built from a common skewness term plus eigenvalue-weighted principal components, which is interpretable and yields closed-form kriging through SUN conditional distributions. On inference, the paper claims a graph attention network plus encoder transformer neural Bayes estimator that outperforms a CNN-based estimator in accuracy and stability.

Load-bearing premise

The construction is a genuine spatial process only if making a spatially correlated Gaussian field non-negative at many locations at once still leaves far-apart locations nearly independent; the paper's proof of that step treats the case of uncorrelated marginals, not the correlated case the model actually uses.

Editorial extensions

If this is right

  • When $\delta_1=\delta_2=0$, $H$ vanishes and GSUN reduces exactly to the Gaussian process, so the model contains Gaussian geostatistics as a limiting case rather than replacing it.
  • At any fixed sample, spatial prediction is available from a closed-form SUN conditional distribution, which yields both kriging means and location-specific prediction intervals without MCMC.
  • The latent truncated process gets its own Matérn range and smoothness parameters, so the model can represent skewness that has a different spatial scale from the underlying Gaussian fluctuations.
  • After training, inference is a single forward pass of the network, avoiding repeated evaluation of high-dimensional Gaussian cumulative distribution functions.
  • The PIT comparisons indicate that GSUN, Gaussian, and Tukey g-and-h fields are distinguishable, so applying GSUN to skewed heavy-tailed data is not just relabeling a Gaussian model.

Reading between the lines

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

  • The same 'Gaussian plus truncated Gaussian' construction could be applied to other covariance families, such as compactly supported or space-time covariances, provided the truncation covariance also decays; the paper does not explore this extension.
  • The diagonal $H$ built from eigenvalue-weighted principal components is a specific modeling choice; one could test whether a full, data-driven $H$ improves fit without sacrificing vanishing correlations or interpretability.
  • The reported advantage of the GAT plus encoder estimator over a CNN estimator on irregular grids suggests the architecture may transfer to other non-Gaussian spatial models with intractable likelihoods, such as Tukey g-and-h or skew-t fields; this is not tested in the paper.
  • A cautious reading is that the finite-dimensional distributions depend on the full observation set through $H$, so the vanishing-correlation argument alone does not settle projective consistency; verifying that the distribution of a subset does not change when distant locations are added would determine whether GSUN is a spatial process in the strict sense.
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 / 5 minor

Summary. The paper proposes a generalized unified skew-normal (GSUN) spatial process defined by Z(s) = W(s) + h(s) W+(s), where W is a Gaussian spatial process with Matérn covariance, W+(s) = (U(s) | U(s) > 0) is a truncated Gaussian process, and h(s) is a location-dependent scalar built from eigenvectors of the covariance and correlation matrices at the observed locations. The authors claim that the GSUN is a valid spatial process with vanishing correlations at large distances, that it admits closed-form kriging, and that it can be estimated by a neural Bayes estimator combining graph attention networks and an encoder transformer. The paper includes a re-parameterization of the SUN distribution, simulation studies comparing the proposed estimator with a CNN-based estimator, uncertainty-quantification experiments, PIT comparisons with Gaussian and Tukey g-and-h fields, and an application to Pb-contaminated soil data.

Significance. If the existence and vanishing-correlation claims were established, the GSUN construction would be a valuable way to introduce skewness and tail flexibility into spatial models while retaining the Gaussian process as a limiting case and enabling fast amortized inference. The paper also contains a useful self-contained derivation of the re-parameterized SUN density, and the GAT/transformer neural Bayes estimator with simulation-based uncertainty quantification is a meaningful engineering contribution. However, the central mathematical claim that the construction defines a stochastic process on the spatial domain is not supported by the arguments given, and the main proof offered for vanishing correlations covers only a diagonal covariance case that is not the model used in the paper. Because the process validity is the foundation for all subsequent claims, the contribution is currently not established.

major comments (3)
  1. [§3.1 (definition of Z(s)=W(s)+h(s)W+(s))] The construction does not define a stochastic process because the finite-dimensional distributions are not Kolmogorov-consistent. The scalar h(s_i) is defined through the eigenvalues and eigenvectors of the full covariance matrix Σ(θ1) and correlation matrix C(θ2) evaluated at the n observed locations; adding or removing an observed location changes h(s_i) for the remaining locations, so the marginal law of Z(s_i) is not stable under extension of the index set. Moreover, W+(s) is defined by the joint conditioning event U(s)>0 over all observed locations simultaneously; for non-diagonal C(θ2), the marginal of U(s1) under the n=2 event U(s1)>0, U(s2)>0 differs from the n=1 definition U(s1)|U(s1)>0. Proposition 2 in the Supplementary Materials treats only diagonal covariance and therefore cannot repair this inconsistency.
  2. [§3.1 (vanishing-correlation argument, with Supplementary S.3)] The claimed vanishing of cov{Z(s_i),Z(s_j)} at large distances relies on the unproved assertion that var(U)_{ij} converges to 0 when the off-diagonal entry C(θ2)_{ij} decays, for the joint truncation event U(s)>0 over all n locations. Proposition 2 assumes a diagonal covariance D, not the Matérn correlation matrix C(θ2) used throughout the model. With non-diagonal C, conditioning on positivity of a third location can induce dependence between locations i and j even if C(θ2)_{ij}=0, so the limiting behavior requires an argument controlling all remaining conditioning coordinates; no such argument is provided.
  3. [§3.2 (conditional distribution for kriging)] The kriging formulas are not usable as written. In the display for the conditional SUN distribution, ξ_{1·2} involves H_{1.2} and γ_{1·2}, which are defined only later in the same display and which themselves depend on the conditioning value z2 through τ_{1·2}; the notation H_{1.2} versus H_{1·2} is also inconsistent. Because the conditional parameters of a SUN distribution depend on z2 in a nontrivial way, the displayed expressions cannot be directly applied to compute conditional means or prediction intervals. This undermines the claimed closed-form spatial interpolation unless the formulas are corrected and verified.
minor comments (5)
  1. [§3.1 (definition of H)] The formula for H is ambiguous: the summands λ_{i,Σ(θ1)} P_{i,Σ(θ1)} + λ_{i,C(θ2)} P_{i,C(θ2)} mix scalars and vectors inside a diag operation, and the intended normalization by 1/(2n) is not explained.
  2. [§5.3.1 and Eq. (4)] The attention normalization in Eq. (4) sums over k∈N_i, but the updated representation aggregates over N_i∪{i}; the treatment of self-connections is inconsistent unless α_{ii} is defined separately.
  3. [Supplementary S.4] The sentence 'Z∼SUN_{n,n}(0, Σ(θ1), H, 0, Σ(θ2))∈ R^{n×n}' appears to contain typographical errors: Z is a vector, and the latent correlation matrix should be C(θ2), not Σ(θ2).
  4. [Figure 6 caption] The parameter vector in the Figure 6 caption contains a malformed symbol '˘3'; this should presumably be −3.
  5. [Throughout] There are repeated typos such as 'ammotized' (Sections 6.1 and S.4) and 'propogation' (Section 5.3.2) that should be corrected in a revision.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the SUN re-parameterization is proved in Section S.1, the neural Bayes estimator is standard amortized inference trained on simulations from the same model, and no fitted constant is relabeled as a prediction.

full rationale

The derivation chain is largely self-contained. The re-parameterized SUN density in Section 2.2 is justified by an explicit proof in Section S.1 that equates the convolution construction to the original Arellano-Valle-Azzalini density via a marginal-conditional Gaussian identity; the self-citation to Arellano-Valle and Genton (2025) is therefore not the load-bearing evidence. The GSUN construction in Section 3.1 specifies H as a deterministic function of the covariance parameters rather than as a fitted parameter, and the paper does not present an independently fitted quantity as a prediction. The neural Bayes estimator is trained and evaluated on simulations from the same generative model; this is the standard amortized-inference setup and is not a circular validation. The main claims that involve self-citations, such as the SUN conditioning formulas used for kriging in Section 3.2, are standard properties with independent derivations in the literature. A genuine mathematical concern exists, namely that the process-validity argument relies on Proposition 2, which proves independence only for diagonal truncation covariance, while the GSUN latent process uses non-diagonal Matérn C(theta2), and the observation-set dependence of H raises Kolmogorov-consistency questions. However, this is a correctness and validity gap rather than a circular reduction of the paper's output to its input.

Assumptions & free parameters 2 free parameters · 7 assumptions · 1 invented entities

The GSUN construction depends on the Matérn model, the equivalence of the SUN re-parameterization, and the unproved assumption that joint truncation of a non-diagonal Gaussian vector preserves vanishing correlations. The model also introduces an unobserved latent truncated Gaussian process. The neural Bayes estimator adds training hyperparameters but these do not affect the theoretical validity claim.

free parameters (2)
  • Eigenvector weighting normalization in H (factor 1/(2n)) = 1/(2n)
    Chosen by hand in Section 3.1 to define h(s) from PCA of the covariance and correlation matrices; it controls how delta2 scales with sample size and is not derived from first principles.
  • Graph radius R = 0.34
    Chosen in Section 5.3.1 so that three GAT layers cover the study region; a hyperparameter affecting the estimator, not the process.
assumptions (7)
  • domain assumption Matérn covariance and correlation functions are valid and have correlations decaying to zero with distance.
    Section 3.1 uses Matérn Sigma(theta1) and C(theta2) without proof; standard in geostatistics but an assumption about the true dependence structure.
  • standard math The re-parameterized SUN distribution (Definition 1) is equivalent to the original SUN distribution.
    Proved in Supplementary S.1 using Gaussian identities; relies on the marginal-conditional decomposition of a multivariate normal density.
  • standard math For a Gaussian vector with diagonal covariance, componentwise truncation factorizes into independent truncated normals.
    Proposition 2 in Supplementary S.3, true for diagonal D.
  • ad hoc to paper For a Gaussian vector with non-diagonal covariance, joint truncation preserves vanishing cross-covariance as off-diagonal entries decay to zero.
    Section 3.1 invokes this to claim var(U)_ij -> 0 at large distances, but Proposition 2 does not apply to non-diagonal C(theta2) with a joint conditioning event; the claim is unproved.
  • ad hoc to paper The finite-dimensional distributions of the GSUN process are consistent and define a stochastic process on the spatial domain.
    Section 3.1 calls GSUN a spatial process, but H and the joint truncation depend on the full set of observed locations; Kolmogorov consistency is not established.
  • standard math The SUN conditional distribution formula used for kriging is correct.
    Taken from Arellano-Valle and Genton (2010) and re-expressed in the new parameterization; accepted in the skew-elliptical literature.
  • domain assumption A sufficiently large and deep neural network can approximate the Bayes estimator arbitrarily well.
    Section 4.2 invokes universal approximation theorems (Hornik et al., 1989; Zhou, 2020); standard but heuristic for finite training.
invented entities (1)
  • Latent truncated Gaussian spatial process W+(s)
    purpose: Introduces skewness into the observed process; Z = W + h W+.
    An unobserved latent process whose existence on a continuum is not established; its only handle is the resulting marginal distribution, so no independent falsifiable prediction outside the model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Generalized Unified Skew-Normal Process with Neural Bayes Inference." pith.science (2026). https://pith.science/paper/I7C3JVI7

@misc{pith2026241117400,
  author       = {Pith},
  title        = {Pith review of: A Generalized Unified Skew-Normal Process with Neural Bayes Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I7C3JVI7}},
  note         = {Machine review of arXiv:2411.17400}
}
read the original abstract

In recent decades, statisticians have been increasingly encountering spatial data that exhibit non-Gaussian behaviors such as asymmetry and heavy-tailedness. As a result, the assumptions of symmetry and fixed tail weight in Gaussian processes have become restrictive and may fail to capture the intrinsic properties of the data. To address the limitations of the Gaussian models, a variety of skewed models has been proposed, of which the popularity has grown rapidly. These skewed models introduce parameters that govern skewness and tail weight. Among various proposals in the literature, unified skewed distributions, such as the Unified Skew-Normal (SUN), have received considerable attention. In this work, we revisit a more concise and intepretable re-parameterization of the SUN distribution and apply the distribution to random fields by constructing a generalized unified skew-normal (GSUN) spatial process. We demonstrate that the GSUN is a valid spatial process by showing its vanishing correlation in large distances and provide the corresponding spatial interpolation method. In addition, we develop an inference mechanism for the GSUN process using the concept of neural Bayes estimators with deep graphical attention networks (GATs) and encoder transformer. We show the superiority of our proposed estimator over the conventional CNN-based architectures regarding stability and accuracy by means of a simulation study and application to Pb-contaminated soil data. Furthermore, we show that the GSUN process is different from the conventional Gaussian processes and Tukey g-and-h processes, through the probability integral transform (PIT).

Figures

Figures reproduced from arXiv: 2411.17400 by the authors.

Figure 1
Figure 1. Graphical representation of the spatial data with undirected edges. The circle in black denotes [PITH_FULL_IMAGE:figures/full_fig_p018_1.png] view at source ↗
Figure 2
Figure 2. GAT- and Encoder-based neural Bayes estimator for the GSUN spatial process. [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Hyper-parameter tuning for the proposed neural Bayes estimator. The red line denotes the [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The boxplots of the parameter estimates obtained from the proposed neural Bayes estimator [PITH_FULL_IMAGE:figures/full_fig_p022_4.png]
Figure 5
Figure 5. Figure 5: demonstrates that the empirical distributions of Θˆ k (z˜n,β) becomes increasingly similar E−400 A−400 E−900 A−900 E−1600 A−1600 σ2 −2 0 2 4 6 E−400 A−400 E−900 A−900 E−1600 A−1600 β1 0.00 0.02 0.04 0.06 E−400 A−400 E−900 A−900 E−1600 A−1600 ν1 0.0 0.5 1.0 1.5 E−400 A−…
Figure 6
Figure 6. Figure 6: The PIT results for a sample realization of the GSUN, Gaussian, Tukey [PITH_FULL_IMAGE:figures/full_fig_p026_6.png]
Figure 7
Figure 7. Figure 7: The PIT results for the 117 observations of Pb-contaminated areas in soils of a region of north [PITH_FULL_IMAGE:figures/full_fig_p028_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 39 canonical work pages

  1. [1]

    G., and Keyes, D

    Abdulah, S., Ltaief, H., Sun, Y., Genton, M. G., and Keyes, D. E. (2018). ExaGeoStat : A high performance unified software for geostatistics on manycore systems. IEEE Transactions on Parallel and Distributed Systems , 29(12):2771--2784

    work page 2018

  2. [2]

    and Naveau, P

    Allard, D. and Naveau, P. (2007). A new spatial skew-normal random field model. Communications in Statistics—Theory and Methods , 36(9):1821--1834

    work page 2007

  3. [3]

    Arellano-Valle, R. B. and Azzalini, A. (2006). On the unification of families of skew-normal distributions. Scandinavian Journal of Statistics , 33(3):561--574

    work page 2006

  4. [4]

    Arellano-Valle, R. B. and Azzalini, A. (2022). Some properties of the unified skew-normal distribution. Statistical Papers , 63:461–487. Correction in 2023 at https://doi.org/10.1007/s00362-023-01412-5

    work page doi:10.1007/s00362-023-01412-5 2022

  5. [5]

    Arellano-Valle, R. B. and Genton, M. G. (2010). Multivariate unified skew-elliptical distributions. Chilean Journal of Statistics , 1(1):17--33

    work page 2010

  6. [6]

    Arellano-Valle, R. B. and Genton, M. G. (2025). Multivariate Statistics Beyond Normality . Chapman and Hall/CRC, in preparation

    work page 2025

  7. [7]

    Azzalini, A. A. (2023). The R package sn : The skew-normal and related distributions such as the skew- t and the SUN (version 2.1.1). Universit\`a degli Studi di Padova, Italia. Home page: http://azzalini.stat.unipd.it/SN/

    work page 2023

  8. [8]

    Ba, J. L. (2016). Layer normalization. arXiv preprint arXiv:1607.06450

    arXiv 2016

Show all 48 references
  1. [9]

    B., and Morales-O \ n ate, V

    Bevilacqua, M., Caama \ n o-Carrillo, C., Arellano-Valle, R. B., and Morales-O \ n ate, V. (2021). Non-gaussian geostatistical modeling using (skew) t processes. Scandinavian Journal of Statistics , 48(1):212--245

    2021

  2. [10]

    P., Selman, B., and Weinberger, K

    Bjorck, N., Gomes, C. P., Selman, B., and Weinberger, K. Q. (2018). Understanding batch normalization. Advances in Neural Information Processing Systems , 31

    2018

  3. [11]

    Dominguez-Molina, J., Gonzalez-Farias, G., and Gupta, A. (2003). The multivariate closed skew normal distribution. Technical Report 03-12

    2003

  4. [12]

    Genton, M. G. and Zhang, H. (2012). Identifiability problems in some non-gaussian spatial random fields. Chilean Journal of Statistics , 3(2):171--179

    2012

  5. [13]

    and Khosravi, H

    Gholamalinezhad, H. and Khosravi, H. (2020). Pooling methods in deep neural networks, a review. arXiv preprint arXiv:2009.07485

    2020 arXiv

  6. [14]

    S., Riley, P

    Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. (2017). Neural message passing for quantum chemistry. In 34th International Conference on Machine Learning (ICML)

    2017

  7. [15]

    Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep L earning . MIT press

    2016

  8. [16]

    K., Aziz, M

    Gupta, A. K., Aziz, M. A., and Ning, W. (2013). On some properties of the unified skew normal distribution. Journal of Statistical Theory and Practice , 7:480--495

    2013

  9. [17]

    Haran, M. (2011). Gaussian random field models for spatial data. Handbook of Markov Chain Monte Carlo , pages 449--478

    2011

  10. [18]

    Hornik, K., Stinchcombe, M., and White, H. (1989). Multilayer feedforward networks are universal approximators. Neural Networks , 2(5):359--366

    1989

  11. [19]

    and Mallick, B

    Kim, H.-M. and Mallick, B. K. (2004). A bayesian prediction using the skew gaussian distribution. Journal of Statistical Planning and Inference , 120(1-2):85--101

    2004

  12. [20]

    Kipf, T. N. and Welling, M. (2016). Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907

    2016 arXiv

  13. [21]

    Lau, M. M. and Lim, K. H. (2018). Review of adaptive activation function in deep neural network. In 2018 IEEE-EMBS Conference on Biomedical Engineering and Sciences (IECBES) , pages 686--690. IEEE

    2018

  14. [22]

    Lenzi, A., Bessac, J., Rudi, J., and Stein, M. L. (2023). Neural networks for parameter estimation in intractable models. Computational Statistics & Data Analysis , 185:107762

    2023

  15. [23]

    Mahmoudian, B. (2018). On the existence of some skew-gaussian random field models. Statistics & Probability Letters , 137:331--335

    2018

  16. [24]

    and Ferracuti, L

    Minozzo, M. and Ferracuti, L. (2012). On the existence of some skew-normal stationary processes. Chilean Journal of Statistics (ChJS) , 3(2)

    2012

  17. [25]

    G., and Keyes, D

    Mondal, S., Abdulah, S., Ltaief, H., Sun, Y., Genton, M. G., and Keyes, D. E. (2023). Tile low-rank approximations of non-gaussian space and space-time tukey g-and-h random field likelihoods and predictions on large-scale systems. Journal of Parallel and Distributed Computing ...

    2023

  18. [26]

    A., Cire s an, D., Meier, U., Giusti, A., Nagi, F., Schmidhuber, J., and Gambardella, L

    Nagi, J., Ducatelle, F., Di Caro, G. A., Cire s an, D., Meier, U., Giusti, A., Nagi, F., Schmidhuber, J., and Gambardella, L. M. (2011). Max-pooling convolutional neural networks for vision-based hand gesture recognition. In 2011 IEEE I nternational C onference on S ignal and ...

    2011

  19. [27]

    Richards, J., Sainsbury-Dale, M., Zammit-Mangion, A., and Huser, R. (2023). Likelihood-free neural bayes estimators for censored peaks-over-threshold models. arXiv preprint arXiv:2306.15642

    2023 arXiv

  20. [28]

    and Omre, H

    Rimstad, K. and Omre, H. (2014). Skew-gaussian random fields. Spatial Statistics , 10:43--62

    2014

  21. [29]

    Sainsbury-Dale, M., Richards, J., Zammit-Mangion, A., and Huser, R. (2023). Neural bayes estimators for irregular spatial data using graph neural networks. arXiv preprint arXiv:2310.02600

    2023 arXiv

  22. [30]

    Sainsbury-Dale, M., Zammit-Mangion, A., and Huser, R. (2024). Likelihood-free parameter estimation with neural bayes estimators. The American Statistician , 78(1):1--14

    2024

  23. [31]

    Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., and Salakhutdinov, R. (2014). Dropout: a simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research , 15(1):1929--1958

    2014

  24. [32]

    Tagle, F., Castruccio, S., Crippa, P., and Genton, M. G. (2019). A non-gaussian spatio-temporal model for daily wind speeds based on a multi-variate skew-t distribution. Journal of Time Series Analysis , 40(3):312--326

    2019

  25. [33]

    Targ, S., Almeida, D., and Lyman, K. (2016). Resnet in resnet: Generalizing residual architectures. arXiv preprint arXiv:1603.08029

    2016 arXiv

  26. [34]

    Tukey, J. W. et al. (1977). Exploratory D ata A nalysis , volume 2. Springer

    1977

  27. [35]

    Vaswani, A. (2017). Attention is all you need. Advances in Neural Information Processing Systems

    2017

  28. [36]

    Veli c kovi \'c , P., Cucurull, G., Casanova, A., Romero, A., Li \`o , P., and Bengio, Y. (2018). Graph attention networks. In 6th International Conference on Learning Representations (ICLR)

    2018

  29. [37]

    Walchessen, J., Lenzi, A., and Kuusela, M. (2024). Neural likelihood surfaces for spatial processes with computationally intensive or intractable likelihoods. Spatial Statistics , 62:100848

    2024

  30. [38]

    Wang, K., Abdulah, S., Sun, Y., and Genton, M. G. (2023a). Which parameterization of the M at \'e rn covariance function?. Spatial Statistics , 58:100787

    2023

  31. [39]

    B., Azzalini, A., and Genton, M

    Wang, K., Arellano-Valle, R. B., Azzalini, A., and Genton, M. G. (2023b). On the non-identifiability of unified skew-normal distributions. Stat , 12(1):e597

    2023

  32. [40]

    J., Arellano-Valle, R

    Wang, K., Karling, M. J., Arellano-Valle, R. B., and Genton, M. G. (2024). Multivariate unified skew-t distributions and their properties. Journal of Multivariate Analysis , 203:105322

    2024

  33. [41]

    and Manjunath, B

    Wilhelm, S. and Manjunath, B. (2023). tmvtnorm : Truncated Multivariate Normal and Student t Distribution . R package version 1.6

    2023

  34. [42]

    and Genton, M

    Xu, G. and Genton, M. G. (2017). Tukey g-and-h random fields. Journal of the American Statistical Association , 112(519):1236--1249

    2017

  35. [43]

    R., and Smola, A

    Zaheer, M., Kottur, S., Ravanbakhsh, S., Poczos, B., Salakhutdinov, R. R., and Smola, A. J. (2017). Deep sets. Advances in neural information processing systems , 30

    2017

  36. [44]

    Zammit-Mangion, A., Sainsbury-Dale, M., and Huser, R. (2024). Neural methods for amortised parameter inference. arXiv preprint arXiv:2404.12484

    2024 arXiv

  37. [45]

    and Khaledi, M

    Zareifard, H. and Khaledi, M. J. (2013). Non- G aussian modeling of spatial data using scale mixing of a unified skew G aussian process. Journal of Multivariate Analysis , 114:16--28

    2013

  38. [46]

    Zhang, H. (2004). Inconsistent estimation and asymptotically equal interpolations in model-based geostatistics. Journal of the American Statistical Association , 99(465):250--261

    2004

  39. [47]

    and El-Shaarawi, A

    Zhang, H. and El-Shaarawi, A. (2010). On spatial skew-gaussian processes and applications. Environmetrics , 21(1):33--47

    2010

  40. [48]

    Zhou, D.-X. (2020). Universality of deep convolutional neural networks. Applied and Computational Harmonic Analysis , 48(2):787--794

    2020

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.