Pith. sign in

REVIEW 4 major objections 8 minor 28 references

PIGPVAE: Physics-Informed Gaussian Process Variational Autoencoders

T0 review · 4 major / 8 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 By combining a physical decoder with a learned discrepancy term, PIGPVAE generates realistic indoor temperature curves from very limited data and extrapolates to starting temperatures absent from training.

desk verdict A sensible integration of physics decoder and GPVAE discrepancy undermined by a trainable alpha that must collapse to zero; fixable, worth a revision. read the letter →

arxiv 2505.19320 v1 pith:PLPXWC2Z submitted 2025-05-25 stat.ML cs.LG

Michail Spitieris , Massimiliano Ruocco , Abdulmajid Murad , Alessandro Nocente This is my paper · ORCID
classification stat.MLcs.LG
keywords physics-informedmachinelearningvariationalautoencodersGaussianprocesspriorssynthetictimeseriesgenerationlimiteddataindoortemperaturemodeldiscrepancyout-of-distribution
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

PIGPVAE is a generative model for time series that joins a physical decoder, Newton's law of cooling, with a Gaussian-process VAE discrepancy term that learns the dynamics the physical law misses. The paper argues that on small indoor-temperature datasets this hybrid preserves the overall heating and cooling trend while adding realistic variation, and that it outperforms purely data-driven baselines on distribution-matching metrics. Its central and most distinctive claim is that conditioning on the physical model lets the generator produce realistic curves for starting temperatures absent from the training data, something the compared data-driven models cannot do. The model targets settings where data collection is expensive and some physical prior knowledge exists, such as HVAC control and indoor temperature forecasting.

What carries the argument

The load-bearing object is the additive decoder $\hat{x} = f_{\text{phy}}(z_{\text{phy}}, x_c) + f_{\delta}(z_{\delta}, \hat{x}_{\text{phy}})$, where $f_{\text{phy}}$ is the exponential solution of Newton's law, $z_{\text{phy}} = k$ is the latent physical parameter, and $f_{\delta}$ is the GPVAE discrepancy decoder driven by $z_{\delta} \sim \mathcal{GP}(0, K(t,t))$. The ELBO combines the PIVAE objective over the physical latent with the GPVAE objective over the discrepancy latent, and the extra term $\alpha\|x - \hat{x}_{\text{phy}}\|^2$ penalizes deviation between the physical decoder alone and the observed data. This mechanism keeps the interpretable physics while representing unmodeled dynamics, and it provides the conditioning structure that the out-of-distribution experiments rely on.

What would settle it

Train PIGPVAE on synthetic data generated exactly from Newton's law with known k and no unmodeled dynamics, then check whether the learned posterior for k recovers the true value and whether the trained penalty weight remains nonzero; if the discrepancy term absorbs most of the reconstruction or the physical parameter estimate drifts from truth, the claimed balancing mechanism is not doing the work.

Watch

Extended reading notes

Core claim

The paper's central claim is that replacing the VAE decoder with a physics equation plus a learned discrepancy term yields better limited-data synthetic time-series generation than either part alone. In PIGPVAE, the decoder output is the sum $\hat{x} = f_{\text{phy}}(z_{\text{phy}}, x_c) + f_{\delta}(z_{\delta}, \hat{x}_{\text{phy}})$, where $f_{\text{phy}}$ is the closed-form solution $T(t) = (T_0 - T_s)\exp(-k t) + T_s$ of Newton's law, $z_{\text{phy}} = k$ is the latent physical parameter, and $f_{\delta}$ is a GPVAE decoder driven by a Gaussian-process latent variable $z_{\delta} \sim \mathcal{GP}(0, K(t,t))$ that models the missing physics. A regularization term $\alpha\|x - \hat{x}_{\text{phy}}\|^2$ is added to the ELBO to keep the physical part aligned with observations and prevent the discrepancy term from dominating. The paper reports that on the RICO indoor temperature data, PIGPVAE achieves lower Maximum Mean Discrepancy and Marginal Distribution Difference than GPVAE, PIVAE, TimeGAN, and TimeVQVAE in in-distribution generation, and that when all curves with starting temperatures below $20\,^\circ\mathrm{C}$ are removed from training, PIGPVAE, unlike the data-driven baselines, still produces realistic curves in that unseen range.

Load-bearing premise

The model's advertised balance between physics and flexibility depends on the penalty weight staying meaningful during training; the paper specifies no constraint, schedule, or reparameterization that would stop that weight from collapsing to zero and leaving the physics term with no influence.

Editorial extensions

If this is right

  • On the RICO heating and cooling data, PIGPVAE reports lower MMD and MDD scores than GPVAE, PIVAE, TimeGAN, and TimeVQVAE in in-distribution generation, and its generated curves keep the monotone heating and cooling trend rather than reversing direction.
  • Because the physical decoder conditions on starting temperature, PIGPVAE generates realistic curves for starting temperatures below 20°C even when all such curves are removed from training, while fully data-driven baselines struggle in the same setting.
  • The additive decomposition preserves the latent physical parameter k as an interpretable quantity while the GPVAE discrepancy term absorbs unmodeled dynamics, so the model offers a path to controlled, explainable synthetic data generation from limited samples.
  • The paper's own caveats bound the claim: the model requires approximate knowledge of the underlying physics, and the Gaussian-process latent has O(n³) complexity in the sequence length, which becomes prohibitive on long horizons.

Reading between the lines

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

  • If the trainable penalty weight is the real source of the reported balance, then a fixed weighting or a constrained version should perform comparably; comparing trainable α against fixed schedules would test whether the architecture or an unstated optimization detail drives the results.
  • The same physics-plus-discrepancy construction should transfer to other physical laws with closed-form solutions or cheap solvers, such as RC-network thermal models, so the method's scope is not limited to exponential cooling curves.
  • A natural stress test for the extrapolation claim is a systematic scan over starting temperatures: train only on one window and measure how far outside that window the generated curves remain realistic, since the paper demonstrates only a single cutoff.
  • Replacing the exact GP with sparse or Markovian approximations should remove the O(n³) bottleneck while keeping the physics decoder; if the discrepancy term stays small under the approximation, the method would scale to long-horizon time series.
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

4 major / 8 minor

Summary. The paper proposes PIGPVAE, a variational autoencoder whose decoder combines a physical model (Newton's law of cooling) with a latent Gaussian Process (GP) discrepancy term, and adds a regularization penalty intended to balance physical fidelity against the flexibility of the discrepancy model. The method is evaluated on indoor temperature data from the RICO dataset, with claims of state-of-the-art synthetic time-series generation and the ability to generate realistic out-of-distribution samples. The central mechanism for balancing physics and flexibility is a regularization weight α, which the paper describes as trainable and optimized jointly with the rest of the model.

Significance. The core idea of augmenting a physics-based decoder with a latent GP discrepancy term is well motivated and addresses a real problem in low-data synthetic time-series generation. The paper evaluates on a real HVAC dataset, compares against four baselines (GPVAE, PIVAE, TimeVQVAE, TimeGAN), and provides visual and quantitative evidence for both in-distribution and out-of-distribution generation. If the regularization mechanism were well-posed, the model could be a useful contribution. However, as described, the trainable α collapses to zero, and the empirical improvements over the best baseline are very small (e.g., MMD differences of 0.001–0.008) relative to the reported standard deviations. The paper also does not yet make the implementation reproducible. These issues are fixable, but the current version does not fully support the claims.

major comments (4)
  1. [Sec. 3.2, Eq. (11)] The regularization weight α is described as a trainable parameter optimized jointly with the model parameters. Because the penalty term α||x − x_phy||^2 is non-negative, the gradient of the total loss with respect to α is ||x − x_phy||^2 ≥ 0, so unconstrained joint minimization drives α to 0. The paper provides no constraint (e.g., α ≥ α_min), no reparameterization (e.g., softplus or sigmoid), and no schedule. Thus the claimed balance between physical fidelity and discrepancy flexibility is not realized by the stated training procedure. If α is in practice fixed or bounded, the text must say so explicitly; if it is truly trainable, the model reduces to an unregularized physical-decoder GPVAE, which undermines the core innovation.
  2. [Sec. 4.2, Tables 1 and 2] The statistical support for the 'state-of-the-art' claim is not established. The p-values are reported without describing the statistical test, the number of random seeds, or any multiple-comparison correction. Moreover, the reported differences are small relative to the standard deviations; for example, in-distribution heating MMD is 0.0688 (0.0029) for PIGPVAE versus 0.0702 (0.0022) for GPVAE, a difference of 0.0014 that is smaller than either standard error. The claim that PIGPVAE outperforms almost all baselines on nearly all metrics requires confidence intervals on the differences or a properly described significance test.
  3. [Sec. 4.3, Figs. 11–13] The out-of-distribution comparison is not controlled for conditioning information. PIVAE and PIGPVAE are conditioned on the initial temperature T0, while GPVAE, TimeGAN, and TimeVQVAE are not given T0 as an input. The conclusion that fully data-driven models struggle to generate samples outside the observed distribution may therefore be an artifact of withholding the conditioning variable rather than a property of the models. A fairer comparison would include a data-driven baseline that also conditions on T0 (e.g., a conditional GPVAE or conditional GAN), or the paper should argue why such conditioning is not applicable to those baselines.
  4. [Appendix A and Section 4] The implementation is not reproducible from the manuscript. The paper does not report the architecture of the discrepancy decoder fδ, the GP kernel hyperparameters, the optimizer, learning rate, number of epochs, or the exact training procedure for α. The statement that code will be publicly available after the review process is not sufficient for reviewers or readers to verify the results. Please include these details or provide the code with the submission.
minor comments (8)
  1. [Title/header] The title and running header contain a typo: 'PIGPV AE' should be 'PIGPVAE'.
  2. [Secs. 3.1 and 3.2] Section 3.1: 'descibed' should be 'described'; Section 3.2: 'employees' should be 'employs'.
  3. [Eqs. (7), (10), (11), (40)] The conditioning variables xc are introduced in Eq. (7) but are not written as explicit conditioning arguments in the likelihood terms of Eqs. (10), (11), and (40); please make the notation consistent throughout.
  4. [Eq. (10) and Appendix A] The term logZ(x,t) in Eq. (10) is not defined; Appendix A.3 and A.4 use logZ(x). State the definition and whether the normalization constant depends on t.
  5. [Tables 1 and 2] The p-values in Tables 1 and 2 should specify the null hypothesis and the test used, and the number of independent training runs should be reported.
  6. [Sec. 4.1, Eq. (13)] The MMD estimator in Eq. (13) uses a Gaussian kernel with bandwidth σ, but the paper does not state how σ is selected or whether the same value is used for all models.
  7. [Sec. 4.2] The text says 'consisting of 20 data points for heating and 20 data points for cooling' after stating the dataset has 29 and 28 curves; clarify whether these numbers refer to the training or evaluation set.
  8. [Figure 5 caption] The Figure 5 caption contains a typo: 'a nd' should be 'and'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the PIGPVAE objective and results are assembled from standard VAE/GPVAE components and external benchmarks; the self-citation [20] is not load-bearing.

full rationale

The paper's derivation chain is self-contained. Eq. (10) is the standard variational lower bound applied to the factorized prior p(z_phy,z_delta)=p(z_delta|z_phy)p(z_phy), and Eq. (11) adds an explicit penalty alpha||x-x_phy||^2; neither equation defines a target quantity in terms of the same target quantity. The physical decoder (Newton's law, Eq. (6)) is an input prior, not a derived prediction. The SOTA claims in Tables 1-2 are evaluated against external baselines (TimeGAN, TimeVQVAE, GPVAE, PIVAE) on the RICO dataset, so they are not fitted inputs relabeled as predictions. The only self-citation, [20] (Spitieris & Steinsland), appears in related work as one of several references on jointly modeling missing physics and is not used to justify any derivation, uniqueness claim, or experimental choice, hence it is not load-bearing. The manuscript's stated limitations (Sec. 5: need for a physics prior; O(n^3) GP cost) are acknowledged but do not create circularity. A separate correctness risk is noted: Sec. 3.2 says alpha is trainable and optimized jointly, but since the penalty alpha||x-x_phy||^2 is non-negative, unconstrained minimization would drive alpha to 0; this is an optimization/inconsistency issue, not a circularity. Overall, no 'prediction' reduces by construction to an input.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The model's extra structure beyond the physics decoder consists of a GP latent discrepancy and a heuristic penalty. The GP kernel hyperparameters and the prior on k are unspecified, and the penalty weight alpha is described as trainable without a mechanism preventing collapse to zero.

free parameters (4)
  • alpha (regularization weight) = not reported; described as trainable
    Introduced in Eq. 11 to balance physics fidelity versus discrepancy flexibility; if jointly minimized without constraint it collapses to 0.
  • GP kernel hyperparameters (lengthscale, variance) = not reported
    The GP prior z_delta ~ GP(0, K(t,t)) requires kernel parameters; the paper does not specify how they are set or learned.
  • Prior hyperparameters for k (mean, variance) = not reported
    The PIVAE prior p(z_phy) has 'defined mean and variance' in Sec 3.1, but the values and how they are chosen are not given.
  • Encoder and discrepancy decoder network weights = not reported
    Standard NN parameters optimized during training; no architecture or initialization details are provided.
assumptions (4)
  • domain assumption Newton's law of cooling is an adequate approximate generative model for indoor temperature curves
    Used as the PIVAE decoder in Eq. 6; the authors acknowledge it captures only main trends and must be supplemented by a discrepancy term.
  • domain assumption The discrepancy latent z_delta follows a zero-mean Gaussian process with covariance K(t,t) and is conditionally independent of z_phy
    Introduced in Sec 3.2 and the prior decomposition p(z_phy,z_delta)=p(z_delta|z_phy)p(z_phy); the independence is stated, not derived from a physical argument.
  • standard math The GPVAE variational approximation q(x|z)=p(z)q*(z|x)/Z(x) carries over to the joint PIGPVAE model
    Used in Appendix A.4 to derive Eq. 40; the conditions under which this approximation remains valid in the joint model are not established.
  • ad hoc to paper The regularizer alpha ||x - x_phy||^2 can be appended to the ELBO without changing the generative model
    Eq. 11 adds a heuristic penalty with no probabilistic derivation; its treatment as a trainable weight is not justified.
invented entities (1)
  • Latent discrepancy process z_delta
    purpose: Captures temporal dynamics not explained by the physics decoder
    Internal GP latent variable with no falsifiable prediction outside the fitted model; it is a modeling device, not an independently testable entity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PIGPVAE: Physics-Informed Gaussian Process Variational Autoencoders." pith.science (2026). https://pith.science/paper/PLPXWC2Z

@misc{pith2026250519320,
  author       = {Pith},
  title        = {Pith review of: PIGPVAE: Physics-Informed Gaussian Process Variational Autoencoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PLPXWC2Z}},
  note         = {Machine review of arXiv:2505.19320}
}
read the original abstract

Recent advances in generative AI offer promising solutions for synthetic data generation but often rely on large datasets for effective training. To address this limitation, we propose a novel generative model that learns from limited data by incorporating physical constraints to enhance performance. Specifically, we extend the VAE architecture by incorporating physical models in the generative process, enabling it to capture underlying dynamics more effectively. While physical models provide valuable insights, they struggle to capture complex temporal dependencies present in real-world data. To bridge this gap, we introduce a discrepancy term to account for unmodeled dynamics, represented within a latent Gaussian Process VAE (GPVAE). Furthermore, we apply regularization to ensure the generated data aligns closely with observed data, enhancing both the diversity and accuracy of the synthetic samples. The proposed method is applied to indoor temperature data, achieving state-of-the-art performance. Additionally, we demonstrate that PIGPVAE can produce realistic samples beyond the observed distribution, highlighting its robustness and usefulness under distribution shifts.

Figures

Figures reproduced from arXiv: 2505.19320 by the authors.

Figure 1
Figure 1. Standard VAE x z x ′ q(z|x) p(x|z) p(z|t) [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 3
Figure 3. PIVAE x zphy zδ x ′ q(zphy|x) q(zδ|x) p(x|zphy) p(x|zδ) p(zphy) p(zδ|t) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. PIGPVAE model; Reconstructed cooling data. Left plot shows [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figures from the paper (8 more)
Figure 6
Figure 6. Figure 6: Visual comparison of generated vs real data. The top figure shows [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: PCA, t-SNE and densities; cooling data. PIGPVAE Original Generated PIVAE GPVAE TimeGAN TimeVQVAE 10 15 20 25 30 35 Temperature (°C) 0.00 0.05 0.10 0.15 0.20 0.25 Density 10 15 20 25 30 35 40 Temperature (°C) 0.00 0.05 0.10 0.15 0.20 0.25 10 15 20 25 30 Temperature (°C)…
Figure 8
Figure 8. Figure 8: PCA, t-SNE and densities; heating data. 10 [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: PCA, t-SNE and densities. In distribution generation; cooling [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: PCA, t-SNE and densities. In distribution generation; [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Visual comparison of generated vs real data. The top figure shows [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]
Figure 12
Figure 12. Figure 12: PCA, t-SNE and densities. Out of distribution generation; cooling [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]
Figure 13
Figure 13. Figure 13: PCA, t-SNE and densities. Out of distribution generation; heating [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 23 canonical work pages

  1. [21]

    Physics-Integrated Vari- ational Autoencoders for Robust and Interpretable Generative Modeling

    Naoya Takeishi and Alexandros Kalousis. Physics-Integrated Vari- ational Autoencoders for Robust and Interpretable Generative Modeling. November 2021

    work page 2021

  2. [1]

    High-resolution image synthesis with la- tent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with la- tent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    work page 2022

  3. [2]

    Generative adversarial nets.Advances in neural informa- tion processing systems, 27, 2014

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets.Advances in neural informa- tion processing systems, 27, 2014

    work page 2014

  4. [3]

    Auto-encoding variational Bayes

    Diederik P Kingma. Auto-encoding variational Bayes. arXiv preprint arXiv:1312.6114, 2013

    arXiv 2013

  5. [4]

    Normalizing flows for probabilistic modeling and inference.Journal of Machine Learning Research, 22(57):1–64, 2021

    George Papamakarios, Eric Nalisnick, Danilo Jimenez Rezende, Shakir Mohamed, and Balaji Lakshminarayanan. Normalizing flows for probabilistic modeling and inference.Journal of Machine Learning Research, 22(57):1–64, 2021

    2021

  6. [5]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  7. [6]

    Time- series generative adversarial networks.Advances in neural infor- mation processing systems, 32, 2019

    Jinsung Yoon, Daniel Jarrett, and Mihaela Van der Schaar. Time- series generative adversarial networks.Advances in neural infor- mation processing systems, 32, 2019

    work page 2019

  8. [7]

    Vector quantized time series generation with a bidirectional prior model

    Daesoo Lee, Sara Malacarne, and Erlend Aune. Vector quantized time series generation with a bidirectional prior model. In Fran- cisco Ruiz, Jennifer Dy, and Jan-Willem van de Meent, editors, Proceedings of The 26th International Conference on Artificial Intelligence and Statistics, volume 206 ofProceedings of Machine Learning Research, pages 7665–7693. PM...

    work page 2023

Show all 28 references
  1. [8]

    Generative time-series modeling with fourier flows

    Ahmed Alaa, Alex James Chan, and Mihaela van der Schaar. Generative time-series modeling with fourier flows. InInternational Conference on Learning Representations, 2021

    2021

  2. [9]

    Tsgbench: Time series generation benchmark

    Yihao Ang, Qiang Huang, Yifan Bao, Anthony KH Tung, and Zhiyong Huang. Tsgbench: Time series generation benchmark. arXiv preprint arXiv:2309.03755, 2023

    2023 arXiv

  3. [10]

    The Gaussian Process Prior VAE for Interpretable Latent Dynamics from Pixels

    Michael Pearce. The Gaussian Process Prior VAE for Interpretable Latent Dynamics from Pixels. InProceedings of The 2nd Sympo- sium on Advances in Approximate Bayesian Inference, pages 1–12. PMLR, February 2020. ISSN: 2640-3498. 17

    2020

  4. [11]

    Scalable gaussian process variational autoencoders

    Metod Jazbec, Matt Ashman, Vincent Fortuin, Michael Pearce, Stephan Mandt, and Gunnar Rätsch. Scalable gaussian process variational autoencoders. In Arindam Banerjee and Kenji Fuku- mizu, editors,Proceedings of The 24th International Conference on Artificial Intelligence and S...

    2021

  5. [12]

    Fully Bayesian Autoencoders with Latent Sparse Gaus- sian Processes

    Ba-Hien Tran, Babak Shahbaba, Stephan Mandt, and Maurizio Filippone. Fully Bayesian Autoencoders with Latent Sparse Gaus- sian Processes. InProceedings of the 40th International Conference on Machine Learning, pages 34409–34430. PMLR, July 2023. ISSN: 2640-3498

    2023

  6. [14]

    Neural discrete rep- resentation learning.Advances in neural information processing systems, 30, 2017

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete rep- resentation learning.Advances in neural information processing systems, 30, 2017

    2017

  7. [15]

    Gaussian process prior variational autoencoders

    Francesco Paolo Casale, Adrian Dalca, Luca Saglietti, Jennifer Listgarten, and Nicolo Fusi. Gaussian process prior variational autoencoders. Advances in neural information processing systems, 31, 2018

    2018

  8. [16]

    The Gaussian process prior vae for interpretable latent dynamics from pixels

    Michael Pearce. The Gaussian process prior vae for interpretable latent dynamics from pixels. InSymposium on advances in ap- proximate bayesian inference, pages 1–12. PMLR, 2020

    2020

  9. [17]

    Physics-informed machine learning

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

    2021

  10. [18]

    Bayesian calibration of computer models

    Marc C Kennedy and Anthony O’Hagan. Bayesian calibration of computer models. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 63(3):425–464, 2001

    2001

  11. [19]

    Deep Gaussian processes for calibration of computer models (with discussion)

    Sébastien Marmin and Maurizio Filippone. Deep Gaussian processes for calibration of computer models (with discussion). Bayesian Analysis, 17(4):1301–1350, 2022

    2022

  12. [20]

    Bayesian calibration of imperfect computer models using physics-informed priors.Journal of Machine Learning Research, 24(108):1–39, 2023

    Michail Spitieris and Ingelin Steinsland. Bayesian calibration of imperfect computer models using physics-informed priors.Journal of Machine Learning Research, 24(108):1–39, 2023. 18

    2023

  13. [22]

    Physics- informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differ- ential equations

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

    2019

  14. [23]

    The rico dataset: a multivariate hvac indoors and outdoors time-series dataset

    Zachari Thiry, Massimiliano Ruocco, Alessandro Nocente, and Odne Andreas Oksavik. The rico dataset: a multivariate hvac indoors and outdoors time-series dataset. Data in Brief, page 111678, 2025

    2025

  15. [24]

    TSGM: A flexible framework for generative modeling of synthetic time series

    Alexander V Nikitin, Letizia Iannucci, and Samuel Kaski. TSGM: A flexible framework for generative modeling of synthetic time series. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024

    2024

  16. [25]

    A kernel two-sample test.The Journal of Machine Learning Research, 13(1):723–773, 2012

    Arthur Gretton, Karsten M Borgwardt, Malte J Rasch, Bernhard Schölkopf, and Alexander Smola. A kernel two-sample test.The Journal of Machine Learning Research, 13(1):723–773, 2012

    2012

  17. [26]

    Sig-Wasserstein GANs for time series generation

    Hao Ni, Lukasz Szpruch, Marc Sabate-Vidales, Baoren Xiao, Mag- nus Wiese, and Shujian Liao. Sig-Wasserstein GANs for time series generation. InProceedings of the Second ACM International Conference on AI in Finance, pages 1–8, 2021

    2021

  18. [27]

    GP-VAE: Deep Probabilistic Time Series Imputation

    Vincent Fortuin, Dmitry Baranchuk, Gunnar Raetsch, and Stephan Mandt. GP-VAE: Deep Probabilistic Time Series Imputation. In Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, pages 1651–1661. PMLR, June

  19. [28]

    Markovian Gaussian process variational autoencoders

    Harrison Zhu, Carles Balsells-Rodas, and Yingzhen Li. Markovian Gaussian process variational autoencoders. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors,Proceedings of the 40th International Conference on Mac...

    2023

  20. [29]

    Fully Bayesian autoencoders with latent sparse Gaus- sian processes

    Ba-Hien Tran, Babak Shahbaba, Stephan Mandt, and Maurizio Filippone. Fully Bayesian autoencoders with latent sparse Gaus- sian processes. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, 19 editors, Proceedings of the 4...

    2023

Pith tools

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