Pith. sign in

REVIEW 3 major objections 5 minor 45 references

Scalable Gaussian Processes with Latent Kronecker Structure

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Latent Kronecker structure brings exact Gaussian-process inference to datasets with millions of points

desk verdict Solid, clearly-written paper with a correct core identity and honest experiments; the main caveats are that the 'exact GP' framing is scoped to partial grids and the abstract oversells the learning-curve results. read the letter →

arxiv 2506.06895 v2 pith:XQ7V3MQM submitted 2025-06-07 cs.LG stat.ML

Jihao Andreas Lin , Sebastian Ament , Maximilian Balandat , David Eriksson , José Miguel Hernández-Lobato , Eytan Bakshy This is my paper
classification cs.LGstat.ML
keywords GaussianprocessesKroneckerproductmissingdatapartialgridexactinferenceiterativelinearsolverspathwiseconditioningscalability
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

Gaussian processes give calibrated uncertainty estimates, but exact inference costs $O(n^3)$ time and quickly becomes intractable as $n$ grows. This paper proposes Latent Kronecker GP (LKGP): when inputs lie on a partial grid—a subset of a Cartesian product $S\times T$—the observed covariance is written exactly as $P(K_{SS}\otimes K_{TT})P^{T}$, with $P$ projecting the full Kronecker product down to the observed entries. That restores Kronecker-style speedups despite missing values: matrix-vector products take $O(p^2q + pq^2)$ time and $O(p^2+q^2)$ memory instead of $O(n^2)$ in both, so iterative solvers and pathwise conditioning perform exact GP inference at scale. On real-world data with up to five million examples, LKGP matches or beats state-of-the-art sparse and variational Gaussian processes in runtime and predictive log-likelihood.

What carries the argument

The central object is the latent Kronecker factorization $K_{XX} = P(K_{SS}\otimes K_{TT})P^{T}$: the observed $n\times n$ covariance is the projection of the $pq\times pq$ Kronecker product onto the $n$ observed entries of the grid. The workhorse identity is $P(A\otimes B)P^{T}\operatorname{vec}(C) = P\operatorname{vec}(B\operatorname{vec}^{-1}(P^{T}\operatorname{vec}(C))A^{T})$, implemented as reshape, zero-padding, and slice-indexing, which delivers the fast matrix-vector product. This identity is what lets iterative solvers and pathwise-conditioning updates exploit Kronecker structure in the presence of missing values.

What would settle it

On a fixed partial-grid dataset, measure the missing ratio at which LKGP's runtime and memory cross those of an ordinary dense iterative GP; the crossing should occur near $\gamma^{*}_{\mathrm{time}} = 1 - \sqrt{1/p + 1/q}$ and $\gamma^{*}_{\mathrm{mem}} = 1 - \sqrt{1/p^2 + 1/q^2}$. Also, computing $P(K_{SS}\otimes K_{TT})P^{T}v$ via the reshape-and-slice identity and comparing it with the explicit dense matrix-vector product should match to floating-point precision; disagreement would refute the exactness claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that the joint covariance matrix of observations on a partially observed Cartesian grid can be represented without approximation as $K_{XX} = P(K_{SS}\otimes K_{TT})P^{T}$, with $P$ a projection matrix that removes rows and columns corresponding to missing grid cells. This 'latent Kronecker structure' keeps the full-grid kernel matrix $K_{SS}\otimes K_{TT}$ latent and materializes only the observed sub-block on demand. Because the projection does not break fast matrix-vector multiplication—$P$ is applied as zero-padding and index slicing, never built explicitly—iterative linear solvers such as conjugate gradients realize each matrix-vector product in $O(p^2q + pq^2)$ time and $O(p^2+q^2)$ memory instead of $O(n^2)$ in both. Combined with pathwise conditioning, the same machinery yields exact posterior samples. The paper validates the construction empirically and gives break-even missing ratios beyond which dense iterative methods are asymptotically preferable.

Load-bearing premise

The load-bearing premise is that the inputs form a partial grid—every observation can be assigned to a pair $(s,t)$ from two fixed index sets $S$ and $T$—and that the kernel factorizes as a product $k_S(s,s')\,k_T(t,t')$; if either fails, the projection matrix $P$ cannot be built and the speedups vanish.

Editorial extensions

If this is right

  • Exact GP inference with product kernels becomes practical for partial-grid datasets with millions of observations, a regime previously dominated by approximate sparse or variational methods.
  • Because the model is exact, predictive uncertainty is not distorted by inducing-point approximations; the experiments show LKGP achieving the best test negative log-likelihood on learning-curve and climate benchmarks.
  • Proposition 3.1 gives a switching rule: LKGP is asymptotically faster than a dense iterative GP while the missing ratio $\gamma$ stays below $1 - \sqrt{1/p + 1/q}$, and cheaper in memory below $1 - \sqrt{1/p^2 + 1/q^2}$.
  • If the temporal factor is stationary with uniform sampling, the same latent structure can be pushed to quasi-linear time in the number of time steps by exploiting Toeplitz structure, as the paper notes in Section 2.

Reading between the lines

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

  • The padding-and-projection recipe may generalize: any structured kernel with a fast matrix-vector product on a full grid could be extended to partial observations, with a break-even formula like Proposition 3.1 deciding when the extension is worthwhile.
  • For data that do not naturally lie on a grid, the authors suggest constructing an artificial grid by local interpolation; that step would break the exactness guarantee, and how much interpolation degrades predictions is a testable question the paper leaves open.
  • On the learning-curve benchmark LKGP has the best test log-likelihood but not the best root-mean-square error, which suggests the exact GP's value lies in trustworthy uncertainty for downstream decisions such as pruning runs in hyperparameter optimization, rather than in point prediction alone.
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 Latent Kronecker GP (LKGP), a method for exact Gaussian process inference when the training inputs lie on a partial grid, i.e., a subset of a Cartesian product S×T, and the kernel is a product kernel. The central identity K_XX = P(K_SS⊗K_TT)P^T is exact, and matrix-vector multiplication can be performed in O(p²q+pq²) time and O(p²+q²) kernel-matrix storage via zero-padding, reshaping, and slice indexing. The method is combined with iterative linear solvers (conjugate gradients) and pathwise conditioning to perform inference of the exact GP model up to solver tolerance. Proposition 3.1 gives asymptotic break-even missing ratios for time and memory, which are validated empirically in the SARCOS experiment. Experiments on SARCOS, LCBench (up to 104k points), and Nordic climate data (up to 5M points) compare LKGP against iterative methods and sparse/variational GPs, showing large speedups and strong predictive performance, though with some metric-dependent results. The paper is accompanied by public code and a clear limitations section acknowledging the partial-grid and product-kernel requirements.

Significance. If the results hold, LKGP is a valuable contribution: it extends Kronecker-structured GP inference from fully gridded data to partially observed grids without introducing model approximations, and it does so with a simple, exact algebraic identity. The derivation of Proposition 3.1 is parameter-free and provides a testable prediction that matches experiments, which is a notable strength. The method is clearly positioned relative to existing sparse, variational, and structured-kernel approaches, and the experiments are extensive and transparent, with code released. The main caveat is that the method's scope is narrower than the abstract's 'exact GP' phrasing suggests, but this is not a technical error in the derivation. The contribution is incremental in the sense that it combines known components (Kronecker MVM, projections, iterative solvers), yet the combination is non-obvious and practically useful for a common data regime.

major comments (3)
  1. [Abstract and Section 1] The paper repeatedly claims 'inference of exact GPs' or 'facilitates highly scalable inference of exact GP models' without immediate qualification, but the method is exact only when the inputs are a subset of a single Cartesian product S×T and the kernel factorizes as a product kernel. The Limitations section (p. 9) states these conditions clearly, but a reader skimming the abstract or introduction could overestimate the method's applicability. Please qualify the central claim in the abstract and introduction, e.g., 'exact GPs on partial grids with product kernels.'
  2. [Section 3 and Appendix C] The efficient MVM identity is central to the method, but the experiments with LKGP in LCBench and the climate datasets use a rank-100 pivoted Cholesky preconditioner. The paper does not explain how this preconditioner is constructed for the projected matrix P(K_SS⊗K_TT)P^T without materializing the n×n matrix, nor does it state the construction's asymptotic cost. If the preconditioner requires O(n·100²) work or a large number of MVMs, it could dominate the O(p²q+pq²) per-iteration cost and materially affect the reported runtime advantages. Please provide the preconditioner construction algorithm and its cost, or cite a reference.
  3. [Section 4, Table 1] The abstract states the method 'outperforms state-of-the-art sparse and variational GPs,' but in the LCBench experiment LKGP's average test RMSE (2.600) is worse than SVGP (1.657) and CaGP (2.000). While LKGP is best in average test NLL and by a large margin in time, the blanket claim is inaccurate. Please qualify the claim by metric and dataset, e.g., 'outperforms on negative log-likelihood and runtime, with mixed RMSE results on learning-curve prediction.'
minor comments (5)
  1. [Section 3, 'Posterior Samples via Pathwise Conditioning'] In the unprojected pathwise conditioning equation, y is used without explicitly stating that it is the length-pq full-grid output vector; the subsequent projected equation then uses the length-n observed vector. Please clarify the dimensions of all vectors in these equations.
  2. [Section 3, 'Discussion of Computational Benefits'] The sentence 'the asymptotic space complexity is reduced from O(n) to O(p+q)' is unclear: the MVM still requires O(n) storage for the vector, so the reduction applies only to the storage of the kernel values. Please rephrase to specify what is being counted.
  3. [Proposition 3.1 and Appendix A] The break-even derivation equates asymptotic complexities while ignoring constant factors and the O(n) overhead of zero-padding and slice indexing. The empirical validation in Figure 3 partly addresses this, but adding a sentence that these are asymptotic and that constant factors may shift the break-even would be helpful.
  4. [Figure 2] The caption does not state the factorization used for the ten-dimensional synthetic datasets. The text mentions a balanced factorization p=q=√n, but the figure should state this directly for clarity.
  5. [Appendix C, 'Learning Curve Prediction'] The number of training iterations differs across methods (LKGP 100 iterations, SVGP 30 epochs, VNNGP and CaGP 1000 epochs). Since runtimes are wall-clock, the comparison is fair, but a brief note explaining why these iteration counts were chosen would aid reproducibility.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the exactness claim is an algebraic identity with stated assumptions, the break-even points are derived and then empirically validated, and the self-citations are non-load-bearing building blocks.

full rationale

LKGP's central identity K_XX = P(K_SS otimes K_TT)P^T is asserted under explicitly stated assumptions: the inputs live on a partial grid S x T and the kernel factorizes as k((s,t),(s',t')) = k_S(s,s')k_T(t,t'). The projection matrix P is constructed as the selector of observed rows and columns, so the equality is a submatrix relation and not a fitted quantity or a prediction extracted from data. The efficiency claim O(p^2q + pq^2) follows algebraically from the Kronecker MVM identity with projections, and Proposition 3.1's break-even points are derived in Appendix A by equating asymptotic complexities. The paper then validates those break-even points empirically on SARCOS, which is a genuine comparison of independently measured runtime and memory behavior. The projection-plus-Kronecker representation of a partially observed grid is a mathematical construction, not a renamed empirical pattern or a self-definitional prediction. The paper does cite prior work by the authors and their coauthors (Lin et al. 2023, 2024a, 2024b; Maddox et al. 2021), but these citations are used as standard building blocks: stochastic-gradient linear solvers, marginal-likelihood gradient estimates, and pathwise conditioning. The relevant formulas are stated in the text, and none of these citations supplies a uniqueness theorem, an unverified ansatz, or a fitted input that forces the central claim. The limitation that data must live on a partial grid is acknowledged in the Limitations section; that is a scope restriction rather than circularity. Overall, the derivation is self-contained and no 'prediction' reduces by construction to a fitted input.

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

The method rests on the structural assumption that observations form a subset of a Cartesian product of two index sets and that the kernel is a product kernel. No additional free parameters or invented entities are introduced beyond standard GP hyperparameters, and those are optimized rather than used to make the derivation work. The iterative solver tolerance means 'exact' is up to numerical precision.

assumptions (4)
  • domain assumption Data inputs lie on a partial Cartesian grid: every observed x_i can be written as (s,t) with s ∈ S, t ∈ T, and the observed set is a subset of S × T.
    Section 3, 'Dealing with Missing Values': the projection matrix P is constructed from the identity by removing rows corresponding to missing values. Without a known latent grid there is no P.
  • domain assumption The kernel is a product kernel, k((s,t),(s',t')) = k_S(s,s') k_T(t,t').
    Section 3, 'Ordinary Kronecker Structure': this factorization is required for the joint covariance to be K_S ⊗ K_T.
  • domain assumption The iterative linear system solver (CG with relative tolerance 0.01) converges closely enough that the posterior is an accurate approximation to the exact GP posterior.
    Used in all experiments (Appendix C); the claim 'exact GP' is model-exact, not solver-exact.
  • standard math Standard Kronecker algebra identity (A⊗B)vec(C)=vec(BCA^T).
    Used in Section 3 to compute projected MVMs without materializing K_XX.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Scalable Gaussian Processes with Latent Kronecker Structure." pith.science (2026). https://pith.science/paper/XQ7V3MQM

@misc{pith2026250606895,
  author       = {Pith},
  title        = {Pith review of: Scalable Gaussian Processes with Latent Kronecker Structure},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XQ7V3MQM}},
  note         = {Machine review of arXiv:2506.06895}
}
read the original abstract

Applying Gaussian processes (GPs) to very large datasets remains a challenge due to limited computational scalability. Matrix structures, such as the Kronecker product, can accelerate operations significantly, but their application commonly entails approximations or unrealistic assumptions. In particular, the most common path to creating a Kronecker-structured kernel matrix is by evaluating a product kernel on gridded inputs that can be expressed as a Cartesian product. However, this structure is lost if any observation is missing, breaking the Cartesian product structure, which frequently occurs in real-world data such as time series. To address this limitation, we propose leveraging latent Kronecker structure, by expressing the kernel matrix of observed values as the projection of a latent Kronecker product. In combination with iterative linear system solvers and pathwise conditioning, our method facilitates inference of exact GPs while requiring substantially fewer computational resources than standard iterative methods. We demonstrate that our method outperforms state-of-the-art sparse and variational GPs on real-world datasets with up to five million examples, including robotics, automated machine learning, and climate applications.

Figures

Figures reproduced from arXiv: 2506.06895 by the authors.

Figure 1
Figure 1. Illustrative example considering input data points {(s1, t1),(s1, t2),(s2, t1),(s2, t2),(s2, t3)}, consisting of two out of three time steps at spatial location s1 and three out of three time steps at spatial location s2. Due to the missing observation at (s1, t3), the inputs cannot be expressed as a Cartesian product and the joint covariance matrix cannot be directly expressed as a Kronecker product. Latent Kroneck… view at source ↗
Figure 2
Figure 2. Illustration of computational resources used during kernel evaluation and matrix multiplication on ten-dimensional synthetic datasets of different sizes. Without latent Kronecker structure, memory usage escalates quickly and kernel evaluation time dominates matrix multiplication time asymptotically. With latent Kronecker structure, computations can be scaled to several orders of magnitude larger datasets under simil… view at source ↗
Figure 3
Figure 3. Predicting the inverse dynamics of an anthropomorphic robot arm with seven degrees of freedom. Compared to standard iterative methods, leveraging latent Kronecker structure results in significantly lower runtime and memory requirements while maintaining the same predictive performance. The asymptotic break-even point, at which both methods asymptotically require the same amount of time or memory, closely matches the… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Learning curve prediction on the Fashion-MNIST data from the LCBench benchmark (Zimmer et al., 2021). Partially observed learning curves are extrapolated into the future. The predictive mean and two standard deviations of various GP models are visualized. All models ar…
Figure 5
Figure 5. Figure 5: Illustration of daily temperature and precipitation data from the Nordic Gridded Climate Dataset (Tveito et al., 2000; 2005). The heatmaps (left and right) visualize snapshots of a single day and subsampled spatial locations. Every spatial location is associated with i…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 42 canonical work pages

  1. [1]

    W., and O’Neil, M

    Ambikasaran, S., Foreman-Mackey, D., Greengard, L., Hogg, D. W., and O’Neil, M. Fast Direct Methods for Gaussian Processes . IEEE Transactions on Pattern Analysis and Machine Intelligence, 38 0 (2): 0 252--265, 2016

    work page 2016

  2. [2]

    Ament, S. E. and Gomes, C. P. Scalable First-Order Bayesian Optimization via Structured Automatic Differentiation . In International Conference on Machine Learning, 2022

    work page 2022

  3. [3]

    Bauer, M., Van der Wilk, M., and Rasmussen, C. E. Understanding Probabilistic Sparse Gaussian Process Approximations . In Advances in Neural Information Processing Systems, volume 29, 2016

    work page 2016

  4. [4]

    V., Chai, K

    Bonilla, E. V., Chai, K. M. A., and Williams, C. K. I. M ulti-task G aussian P rocess P rediction. In Advances in Neural Information Processing Systems, 2007

    work page 2007

  5. [5]

    Deisenroth, M. P. and Rasmussen, C. E. PILCO: A Model-Based and Data-Efficient Approach to Policy Search . In International Conference on Machine Learning, 2011

    work page 2011

  6. [6]

    Dong, K., Eriksson, D., Nickisch, H., Bindel, D., and Wilson, A. G. Scalable Log Determinants for Gaussian Process Kernel Learning . Advances in Neural Information Processing Systems, 30, 2017

    work page 2017

  7. [7]

    H., and Hutter, F

    Elsken, T., Metzen, J. H., and Hutter, F. Neural Architecture Search: A Survey . Journal of Machine Learning Research, 20 0 (55): 0 1--21, 2019

    work page 2019

  8. [8]

    Frazier , P. I. A Tutorial on Bayesian Optimization . arXiv e-prints, pp.\ arXiv:1807.02811, July 2018

    arXiv 2018

Show all 45 references
  1. [9]

    Q., Bindel, D., and Wilson, A

    Gardner, J., Pleiss, G., Weinberger, K. Q., Bindel, D., and Wilson, A. G. GPyTorch : B lackbox M atrix-matrix G aussian P rocess I nference with GPU A cceleration. In Advances in Neural Information Processing Systems, 2018 a

    2018

  2. [10]

    Product Kernel Interpolation for Scalable Gaussian Processes

    Gardner, J., Pleiss, G., Wu, R., Weinberger, K., and Wilson, A. Product Kernel Interpolation for Scalable Gaussian Processes . In International Conference on Artificial Intelligence and Statistics, pp.\ 1407--1416, 2018 b

    2018

  3. [11]

    Bayesian Optimization

    Garnett, R. Bayesian Optimization . Cambridge University Press, 2023

    2023

  4. [12]

    Golub, G. H. and Van Loan, C. F. Matrix Computations . JHU press, 2013

    2013

  5. [13]

    and S \"a rkk \"a , S

    Hartikainen, J. and S \"a rkk \"a , S. Kalman filtering and smoothing solutions to temporal gaussian process regression models. In 2010 IEEE international workshop on machine learning for signal processing, pp.\ 379--384. IEEE, 2010

    2010

  6. [14]

    Hensman, J., Fusi, N., and Lawrence, N. D. Gaussian Processes for Big Data . In Proceedings of the Twenty-Ninth Conference on Uncertainty in Artificial Intelligence, pp.\ 282--290, 2013

    2013

  7. [15]

    Parametric Gaussian Process Regressors

    Jankowiak, M., Pleiss, G., and Gardner, J. Parametric Gaussian Process Regressors . In International Conference on Machine Learning, volume 119, pp.\ 4702--4712, 2020

    2020

  8. [16]

    A., and Wilson, A

    Kapoor, S., Finzi, M., Wang, K. A., and Wilson, A. G. G. SKIing on Simplices: Kernel Interpolation on the Permutohedral Lattice for Scalable Gaussian Processes . In International Conference on Machine Learning, pp.\ 5279--5289, 2021

    2021

  9. [17]

    B., Knoblauch, J., and Briol, F.-X

    Laplante, W., Altamirano, M., Duncan, A. B., Knoblauch, J., and Briol, F.-X. Robust and conjugate spatio-temporal gaussian processes. In Forty-second International Conference on Machine Learning, 2025. URL https://openreview.net/forum?id=YG84SWm7gn

    2025

  10. [18]

    A., Antorán, J., Padhy, S., Janz, D., Hernández-Lobato, J

    Lin, J. A., Antorán, J., Padhy, S., Janz, D., Hernández-Lobato, J. M., and Terenin, A. S ampling from G aussian P rocess P osteriors using S tochastic G radient D escent. In Advances in Neural Information Processing Systems, 2023

    2023

  11. [19]

    A., Padhy, S., Antorán, J., Tripp, A., Terenin, A., Szepesvári, C., Hernández-Lobato, J

    Lin, J. A., Padhy, S., Antorán, J., Tripp, A., Terenin, A., Szepesvári, C., Hernández-Lobato, J. M., and Janz, D. S tochastic G radient D escent for G aussian P rocesses D one R ight. In International Conference on Learning Representations, 2024 a

    2024

  12. [20]

    A., Padhy, S., Mlodozeniec, B., Antorán, J., and Hernández-Lobato, J

    Lin, J. A., Padhy, S., Mlodozeniec, B., Antorán, J., and Hernández-Lobato, J. M. I mproving L inear S ystem S olvers for H yperparameter O ptimisation in I terative G aussian P rocesses. In Advances in Neural Information Processing Systems, 2024 b

    2024

  13. [21]

    J., Balandat, M., Wilson, A

    Maddox, W. J., Balandat, M., Wilson, A. G., and Bakshy, E. B ayesian O ptimization with H igh- D imensional O utputs. In Advances in Neural Information Processing Systems, 2021

    2021

  14. [22]

    Computation-aware kalman filtering and smoothing

    Pf \"o rtner, M., Wenger, J., Cockayne, J., and Hennig, P. Computation-aware kalman filtering and smoothing. In Li, Y., Mandt, S., Agrawal, S., and Khan, E. (eds.), Proceedings of The 28th International Conference on Artificial Intelligence and Statistics, volume 258 of Procee...

    2025

  15. [23]

    and Rasmussen, C

    Quinonero-Candela, J. and Rasmussen, C. E. A unifying view of sparse approximate Gaussian process regression . Journal of Machine Learning Research, 6: 0 1939--1959, 2005

    1939

  16. [24]

    Rasmussen, C. E. and Williams, C. K. I. G aussian P rocesses for M achine L earning . MIT Press, 2006

    2006

  17. [25]

    Bayesian Active Learning with Fully Bayesian Gaussian Processes

    Riis, C., Antunes, F., H \"u ttel, F., Lima Azevedo, C., and Pereira, F. Bayesian Active Learning with Fully Bayesian Gaussian Processes . In Advances in Neural Information Processing Systems, 2022

    2022

  18. [26]

    P., Ament, S

    Ryan, J. P., Ament, S. E., Gomes, C. P., and Damle, A. The Fast Kernel Transform . In International Conference on Artificial Intelligence and Statistics, 2022

    2022

  19. [27]

    Scalable Inference for Structured Gaussian Process Models

    Saat c i, Y. Scalable Inference for Structured Gaussian Process Models . PhD thesis, University of Cambridge, 2012

    2012

  20. [28]

    and Svensson, L

    S\"arkk\"a, S. and Svensson, L. Bayesian Filtering and Smoothing. Cambridge University Press, 2nd edition, 2023. ISBN 9781108926645

    2023

  21. [29]

    Spatiotemporal learning via infinite-dimensional bayesian filtering and smoothing: A look at gaussian process regression through kalman filtering

    Sarkka, S., Solin, A., and Hartikainen, J. Spatiotemporal learning via infinite-dimensional bayesian filtering and smoothing: A look at gaussian process regression through kalman filtering. IEEE Signal Processing Magazine, 30 0 (4): 0 51--61, 2013. doi:10.1109/MSP.2013.2246292

    2013

  22. [30]

    and Ghahramani, Z

    Snelson, E. and Ghahramani, Z. Sparse Gaussian Processes using Pseudo-inputs . In Weiss, Y., Sch\" o lkopf, B., and Platt, J. (eds.), Advances in Neural Information Processing Systems, volume 18. MIT Press, 2005

    2005

  23. [31]

    M., Lawrence, N., and Borgwardt, K

    Stegle, O., Lippert, C., Mooij, J. M., Lawrence, N., and Borgwardt, K. E fficient inference in matrix-variate G aussian models with iid observation noise. In Advances in Neural Information Processing Systems, 2011

    2011

  24. [32]

    Variational learning of inducing variables in sparse Gaussian processes

    Titsias, M. Variational learning of inducing variables in sparse Gaussian processes . In International Conference on Artificial Intelligence and Statistics, pp.\ 567--574, 2009

    2009

  25. [33]

    E., Førland, E

    Tveito, O. E., Førland, E. J., Heino, R., Hanssen-Bauer, I., Alexandersson, H., Dahlström, B., Drebs, A., Kern-Hansen, C., Jónsson, T., Vaarby-Laursen, E., and Westman, E. Nordic Temperature Maps . DNMI Klima 9/00 KLIMA., 2000

    2000

  26. [34]

    E., Bjørdal, I., Skjelvåg, A

    Tveito, O. E., Bjørdal, I., Skjelvåg, A. O., and Aune, B. A GIS-based agro-ecoglogical decision system based on gridded climatology . Metoeorl. Appl., 12: 0 57--68, 2005

    2005

  27. [35]

    A., Pleiss, G., Gardner, J

    Wang, K. A., Pleiss, G., Gardner, J. R., Tyree, S., Weinberger, K. Q., and Wilson, A. G. E xact G aussian P rocesses on a M illion D ata P oints. In Advances in Neural Information Processing Systems, 2019

    2019

  28. [36]

    R., Pleiss, G., and Cunningham, J

    Wenger, J., Wu, K., Hennig, P., Gardner, J. R., Pleiss, G., and Cunningham, J. P. Computation-Aware Gaussian Processes: Model Selection And Linear-Time Inference . In Advances in Neural Information Processing Systems, 2024

    2024

  29. [37]

    and Nickisch, H

    Wilson, A. and Nickisch, H. Kernel Interpolation for Scalable Structured Gaussian Processes (KISS-GP) . In International Conference on Machine Learning, pp.\ 1775--1784, 2015

    2015

  30. [38]

    G., Gilboa, E., Nehorai, A., and Cunningham, J

    Wilson, A. G., Gilboa, E., Nehorai, A., and Cunningham, J. P. Fast Kernel Learning for Multidimensional Pattern Extrapolation . In Advances in Neural Information Processing Systems, volume 27, 2014

    2014

  31. [39]

    T., Borovitskiy, V., Terenin, A., Mostowsky, P., and Deisenroth, M

    Wilson, J. T., Borovitskiy, V., Terenin, A., Mostowsky, P., and Deisenroth, M. P. E fficiently S ampling F unctions from G aussian P rocess P osteriors. In International Conference on Machine Learning, 2020

    2020

  32. [40]

    T., Borovitskiy, V., Terenin, A., Mostowsky, P., and Deisenroth, M

    Wilson, J. T., Borovitskiy, V., Terenin, A., Mostowsky, P., and Deisenroth, M. P. P athwise C onditioning of G aussian P rocesses. Journal of Machine Learning Research, 22, 2021

    2021

  33. [41]

    Wu, K., Wenger, J., Jones, H., Pleiss, G., and Gardner, J. R. L arge- S cale G aussian P rocesses via A lternating P rojection. In International Conference on Artificial Intelligence and Statistics, 2024

    2024

  34. [42]

    Wu, L., Pleiss, G., and Cunningham, J. P. Variational Nearest Neighbor Gaussian Process . In International Conference on Machine Learning, 2022

    2022

  35. [43]

    Zhe, S., Xing, W., and Kirby, R. M. S calable H igh- O rder G aussian P rocess R egression. In International Conference on Artificial Intelligence and Statistics, 2019

    2019

  36. [44]

    A uto- P y T orch T abular: M ulti- F idelity M eta L earning for E fficient and R obust A uto DL

    Zimmer, L., Lindauer, M., and Hutter, F. A uto- P y T orch T abular: M ulti- F idelity M eta L earning for E fficient and R obust A uto DL . IEEE Transactions on Pattern Analysis and Machine Intelligence, 43 0 (9): 0 3079 -- 3090, 2021

    2021

  37. [45]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.