arxiv: 2604.11237 · v1 · submitted 2026-04-13 · 💻 cs.CE

Recognition: unknown

A Physics-Aware Variational Graph Autoencoder for Joint Modal Identification with Uncertainty Quantification

Bhargav Nath , Mehulkumar Lakhadive , Anshu Sharma , Basuraj Bhowmik

Authors on Pith no claims yet

Pith reviewed 2026-05-10 15:21 UTC · model grok-4.3

classification 💻 cs.CE

keywords modal identificationgraph neural networksvariational autoencoderuncertainty quantificationtruss structurespower spectral densitystructural dynamicsphysics-informed learning

0 comments

The pith

A variational graph autoencoder jointly recovers natural frequencies, damping ratios, and full mode shapes from frequency-domain truss data while quantifying uncertainty.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a graph neural network model that treats truss structures as graphs whose nodes carry power spectral density values and geometric details. A residual variational encoder learns a latent space that feeds both global predictions of modal frequencies and damping via evidential regression and a separate decoder that reconstructs spatially distributed mode shapes. Orthogonality regularization and mode-shape reconstruction losses are added to keep the outputs physically consistent. The approach is tested on simulated truss populations across different noise levels and sensor densities, showing stable accuracy and uncertainty calibration where traditional methods often degrade.

Core claim

By representing each truss as a graph with PSD and geometry node attributes plus connectivity edges, a residual GraphSAGE variational autoencoder with attention pooling can simultaneously predict natural frequencies and damping ratios through evidential regression and reconstruct full-field mode shapes via a node-level decoder that fuses global latent vectors with local features; physical consistency is enforced by mode-shape reconstruction and orthogonality regularization, yielding calibrated uncertainty estimates on numerically generated data under varying signal-to-noise ratios and sensor availability.

What carries the argument

UResVGAE: a residual GraphSAGE variational encoder on PSD-geometry graphs with attention-driven pooling, evidential regression head for global modal parameters, node decoder for mode shapes, and orthogonality plus reconstruction regularizers.

If this is right

Natural frequencies and damping ratios are predicted accurately even when measurement noise is high and sensors are sparse.
Full-field mode shapes are recovered with high modal assurance criterion values from the same latent representation used for global parameters.
Predictive uncertainty remains consistent with empirical coverage across different signal-to-noise conditions.
Global and node-level modal information are obtained jointly without separate post-processing steps.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The graph formulation could be tested on other linear structures such as beams or frames to check whether the same encoder-decoder structure transfers without retraining from scratch.
Embedding the trained model in a digital-twin loop would allow continuous updating of modal estimates as new frequency-domain data arrive.
Because uncertainty is produced at both global and local levels, the outputs could directly feed probabilistic structural-health-monitoring algorithms that decide inspection priorities.

Load-bearing premise

Encoding truss structures as graphs with PSD and geometric node attributes, then applying mode-shape reconstruction and orthogonality regularization inside the variational graph autoencoder, will produce physically consistent outputs and reliable uncertainty estimates.

What would settle it

On independent real-world truss vibration experiments with known modal parameters, the model's uncertainty intervals fail to contain the measured errors at the claimed coverage rate or the reconstructed mode shapes yield modal assurance criterion values below 0.8 under realistic sensor sparsity.

Figures

Figures reproduced from arXiv: 2604.11237 by Anshu Sharma, Basuraj Bhowmik, Bhargav Nath, Mehulkumar Lakhadive.

**Figure 2.** Figure 2: Signed relative error distributions for natural frequency and damping ratio predictions across all test samples. [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗

**Figure 3.** Figure 3: Comparison of true and predicted mode shapes for a representative test sample. The top row shows the [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

**Figure 4.** Figure 4: Distribution of MAC values across all test samples for Modes 1–4. The violin plots illustrate the spread of [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗

**Figure 5.** Figure 5: Predicted-versus-true scatter plots for the global modal parameters over the test set. Panel (a) shows the [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗

**Figure 6.** Figure 6: Quantitative contribution of epistemic and aleatoric uncertainty across the first four modes for frequency and [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗

**Figure 7.** Figure 7: Predictive distributions of modal parameters for representative test samples. The Gaussian curves represent the [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗

**Figure 8.** Figure 8: Calibration performance across confidence levels for frequency ( [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗

**Figure 9.** Figure 9: Distribution of MAC values for Modes 1–4 under varying noise conditions (Clean, 30 dB, 20 dB, and 10 dB [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗

**Figure 10.** Figure 10: Comparison of natural frequency prediction performance between the baseline GNN and the proposed [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗

**Figure 11.** Figure 11: Comparison of predicted mode shapes obtained using the baseline GNN and the proposed UResVGAE [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗

**Figure 12.** Figure 12: Per-mode frequency prediction scatter plots for GNN and UResVGAE across all SNR levels. Each row [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗

**Figure 13.** Figure 13: Per-mode damping prediction scatter plots for GNN and UResVGAE across all SNR levels. Each row [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗

read the original abstract

Reliable modal identification from output-only vibration data remains a challenging problem under measurement noise, sparse sensing, and structural variability. These challenges intensify when global modal quantities and spatially distributed mode shapes must be estimated jointly from frequency-domain data. This work presents a physics-aware variational graph autoencoder, termed UResVGAE, for joint modal identification with uncertainty quantification from power spectral density (PSD) representations of truss structures. The framework represents each structure as a graph in which node attributes encode PSD and geometric information, while edges capture structural connectivity. A residual GraphSAGE-based encoder, attention-driven graph pooling, and a variational latent representation are combined to learn both graph-level and node-level modal information within a single, unified formulation. Natural frequencies and damping ratios are predicted through evidential regression, and full-field mode shapes are reconstructed through a dedicated node-level decoder that fuses global latent information with local graph features. Physical consistency is promoted via mode-shape reconstruction and orthogonality regularisation. The framework is assessed on numerically generated truss populations under varying signal-to-noise ratios and sensor availability. Results demonstrate accurate prediction of natural frequencies, damping ratios, and mode shapes, with high modal assurance criterion values and stable performance under noisy and sparse sensing conditions. Reliability analysis indicates that the predictive uncertainty is broadly consistent with empirical coverage. The proposed framework offers a coherent and physically grounded graph-based route for joint modal identification with calibrated uncertainty from frequency-domain structural response data.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper builds a variational graph autoencoder that jointly pulls frequencies, damping, and mode shapes from PSD graphs of trusses while adding uncertainty estimates, but physical consistency rests on soft regularization and all tests stay on simulated data.

read the letter

The main advance is the UResVGAE setup: residual GraphSAGE encoder, attention pooling, variational latent space, evidential regression for global modal parameters, and a node decoder for mode shapes, all tied together with orthogonality regularization on truss graphs whose nodes carry PSD and geometry features. This gives a single model for the joint task instead of separate identification steps, and the uncertainty output is checked for coverage on the simulated cases. That combination is new enough on the abstract description and handles the graph structure of connectivity in a natural way for this domain. Performance looks stable across the reported noise and sensor-density variations in the numerical truss populations. The soft spots are clear and central. Consistency is enforced only through reconstruction plus an orthogonality penalty in the loss; there is no hard constraint from the eigenvalue problem or the underlying finite-element operators. On sparse or noisy inputs the learned embeddings can meet the penalty without producing modes whose frequencies match the actual truss geometry and mass/stiffness. All results come from numerically generated data whose generation assumptions align with the model, so we lack external benchmarks or real-structure tests. The abstract states high MAC values and consistent uncertainty but gives no numbers, baselines, or error bars, which makes it hard to judge how much better this is than prior methods. This is for structural-health-monitoring researchers who already work with graph representations of vibration data and want to explore variational or evidential extensions. A reader interested in physics-informed GNNs might borrow pieces of the architecture. For anyone needing tools that work on measured structures, the current evidence is too preliminary. Send it for peer review. The problem is real, the architecture is coherent, and the gaps are fixable with more validation and tighter physical constraints.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes UResVGAE, a variational graph autoencoder for joint modal identification and uncertainty quantification from frequency-domain PSD data of truss structures. Structures are represented as graphs with PSD and geometric node attributes plus connectivity edges; a residual GraphSAGE encoder, attention pooling, variational latent space, evidential regression for frequencies/damping, and node-level mode-shape decoder are combined, with physical consistency promoted via reconstruction loss and orthogonality regularization. The framework is evaluated on numerically generated truss populations under varying SNR and sensor sparsity, claiming accurate predictions, high MAC values, and consistent uncertainty coverage.

Significance. If the quantitative claims hold and the soft constraints prove sufficient, the work would offer a unified graph-based approach to joint modal parameter and mode-shape estimation with calibrated uncertainty, which could be useful for structural health monitoring under noise and incomplete sensing. The integration of evidential regression and graph pooling is a constructive element, though its advantage over existing methods remains to be demonstrated with baselines.

major comments (3)

Abstract: the claims of 'accurate prediction of natural frequencies, damping ratios, and mode shapes, with high modal assurance criterion values and stable performance' are stated without any numerical metrics, error bars, baseline comparisons, or tabulated results, preventing assessment of whether the method improves on conventional output-only techniques such as SSI or frequency-domain decomposition.
Model formulation and loss function: physical consistency is enforced only through a reconstruction term plus a soft orthogonality penalty; no hard constraint is imposed from the discrete eigenvalue problem (K − ω²M)φ = 0 or the underlying finite-element operators, so the learned embeddings may satisfy the regularizer while producing mode shapes whose associated frequencies are inconsistent with the truss geometry and connectivity, particularly under sparse sensors or high noise.
Evaluation section: all reported results derive from numerically generated truss populations whose generation assumptions are not independently verified or cross-checked against experimental data; this creates circularity in which performance depends on the fitted network parameters rather than external benchmarks, undermining generalization claims.

minor comments (2)

Abstract: the acronym UResVGAE is introduced without expansion; provide the full name on first use.
Notation: ensure consistent symbols for PSD matrices, mode-shape vectors, and damping ratios across text, equations, and figures.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and constructive comments on our manuscript. We provide point-by-point responses to the major comments below, indicating where revisions will be made to address the concerns.

read point-by-point responses

Referee: Abstract: the claims of 'accurate prediction of natural frequencies, damping ratios, and mode shapes, with high modal assurance criterion values and stable performance' are stated without any numerical metrics, error bars, baseline comparisons, or tabulated results, preventing assessment of whether the method improves on conventional output-only techniques such as SSI or frequency-domain decomposition.

Authors: We agree with the referee that including quantitative metrics would improve the abstract. In the revised manuscript, we will update the abstract to include key numerical results from our experiments, such as average errors for frequencies and damping, MAC values, and performance under different SNR levels. This will provide a clearer assessment of the method's performance. revision: yes
Referee: Model formulation and loss function: physical consistency is enforced only through a reconstruction term plus a soft orthogonality penalty; no hard constraint is imposed from the discrete eigenvalue problem (K − ω²M)φ = 0 or the underlying finite-element operators, so the learned embeddings may satisfy the regularizer while producing mode shapes whose associated frequencies are inconsistent with the truss geometry and connectivity, particularly under sparse sensors or high noise.

Authors: We acknowledge that our method uses soft constraints through the reconstruction loss and orthogonality penalty rather than hard constraints derived from the eigenvalue problem. This choice is deliberate to enable operation on PSD data alone without requiring the full finite element model, which is often unavailable. The results demonstrate that the predicted mode shapes achieve high MAC values and the frequencies are consistent within the uncertainty bounds. In the revision, we will add a dedicated paragraph discussing the implications of soft versus hard constraints and include sensitivity analysis to show robustness under sparse and noisy conditions. revision: partial
Referee: Evaluation section: all reported results derive from numerically generated truss populations whose generation assumptions are not independently verified or cross-checked against experimental data; this creates circularity in which performance depends on the fitted network parameters rather than external benchmarks, undermining generalization claims.

Authors: The evaluation is indeed performed on synthetic data from finite element simulations, which allows for precise control over noise and sensor placement to test the method's limits. The generation process follows standard modal analysis procedures, and we will provide additional verification details in the revised manuscript to show that the PSD data corresponds to the expected modal parameters. The model learns solely from the graph representation of PSD and geometry, without direct access to the eigenvalue solver. We will clarify this in the text to address potential circularity concerns. Experimental data validation is an important next step but lies beyond the current scope of this numerical investigation. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical ML framework with independent test evaluation

full rationale

The paper proposes an architecture (residual GraphSAGE encoder, attention pooling, variational latent space, evidential regression heads, node decoder, reconstruction loss plus orthogonality regularizer) trained end-to-end on synthetic truss graphs. Performance metrics (frequency/damping accuracy, MAC values, uncertainty calibration) are reported on held-out numerically generated test structures under controlled SNR and sensor sparsity. These results are produced by gradient descent on the composite loss; they do not reduce by algebraic identity, parameter renaming, or self-citation to the training inputs. No uniqueness theorem, ansatz, or hard eigenvalue constraint is invoked whose justification collapses to prior work by the same authors. The method is therefore a standard data-driven model whose claims rest on empirical generalization rather than definitional equivalence.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

Review based on abstract only; specific free parameters, training details, and exact regularization formulations are not available. The framework rests on standard graph neural network assumptions plus domain assumptions about truss graph representations.

axioms (2)

domain assumption Truss structures can be faithfully represented as graphs with PSD and geometric node features and connectivity edges.
Invoked in the framework description for encoding structural information.
domain assumption Mode-shape reconstruction and orthogonality regularization promote physical consistency without introducing systematic bias.
Stated as the mechanism for physical grounding.

invented entities (1)

UResVGAE no independent evidence
purpose: Joint modal identification with uncertainty quantification from PSD graphs
New model architecture introduced in the paper.

pith-pipeline@v0.9.0 · 5569 in / 1349 out tokens · 57065 ms · 2026-05-10T15:21:42.062839+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

38 extracted references · 5 canonical work pages · 2 internal anchors

[1]

System identification methods for (operational) modal analysis: review and comparison

Edwin Reynders. System identification methods for (operational) modal analysis: review and comparison. Archives of computational methods in engineering, 19(1):51–124, 2012. 21 Physics-Aware U-Res VGAE

2012
[2]

Damage identification and health monitoring of structural and mechanical systems from changes in their vibration characteristics: a literature review

Scott W Doebling, Charles R Farrar, Michael B Prime, and Daniel W Shevitz. Damage identification and health monitoring of structural and mechanical systems from changes in their vibration characteristics: a literature review. 1996

1996
[3]

John Wiley & Sons, 2015

Rune Brincker and Carlos Ventura.Introduction to operational modal analysis. John Wiley & Sons, 2015

2015
[4]

Comparison of response transmissibility and power spectral density transmissibil- ity on operational modal analysis.Mechanical Systems and Signal Processing, 160:107912, 2021

Jie Kang, Hehua Ju, and Li Liu. Comparison of response transmissibility and power spectral density transmissibil- ity on operational modal analysis.Mechanical Systems and Signal Processing, 160:107912, 2021

2021
[5]

A brief introduction to recent developments in population-based structural health monitoring.Frontiers in Built Environment, 6:146, 2020

Keith Worden, Lawrence A Bull, Paul Gardner, Julian Gosliga, Timothy J Rogers, Elizabeth J Cross, Evangelos Papatheou, Weijiang Lin, and Nikolaos Dervilis. A brief introduction to recent developments in population-based structural health monitoring.Frontiers in Built Environment, 6:146, 2020

2020
[6]

Damping estima- tion of a pedestrian footbridge–an enhanced frequency-domain automated approach.Journal of Vibroengineering, 23(1):14–25, 2020

Basuraj Bhowmik, Budhaditya Hazra, Michael O’byrne, Bidisha Ghosh, and Vikram Pakrashi. Damping estima- tion of a pedestrian footbridge–an enhanced frequency-domain automated approach.Journal of Vibroengineering, 23(1):14–25, 2020

2020
[7]

Treatment and effect of noise modelling in bayesian operational modal analysis.Mechanical Systems and Signal Processing, 186:109776, 2023

Xinda Ma, Zuo Zhu, and Siu-Kui Au. Treatment and effect of noise modelling in bayesian operational modal analysis.Mechanical Systems and Signal Processing, 186:109776, 2023

2023
[8]

Wen-Hwa Wu, Sheng-Wei Wang, Chien-Chou Chen, and Gwolong Lai. Modal parameter identification for closely spaced modes of civil structures based on an upgraded stochastic subspace methodology.Structure and Infrastructure Engineering, 15(3):296–313, 2019

2019
[9]

Advancements in online modal identification: A recursive simultaneous diagonalization comprehensive framework for real-time applications.Engineering Structures, 305:117770, 2024

Basuraj Bhowmik. Advancements in online modal identification: A recursive simultaneous diagonalization comprehensive framework for real-time applications.Engineering Structures, 305:117770, 2024

2024
[10]

CRC Press, 2022

Basuraj Bhowmik, Budhaditya Hazra, and Vikram Pakrashi.Real-time structural health monitoring of vibrating systems. CRC Press, 2022

2022
[11]

Fame–ssa: A fast adaptive multivariate decomposition method for unsupervised damage detection in built infrastructures.Journal of Sound and Vibration, 629:119676, 2026

Lakhadive Mehulkumar R, Anshu Sharma, and Basuraj Bhowmik. Fame–ssa: A fast adaptive multivariate decomposition method for unsupervised damage detection in built infrastructures.Journal of Sound and Vibration, 629:119676, 2026

2026
[12]

B Bhowmik, T Tripura, B Hazra, and V Pakrashi. Real time structural modal identification using recursive canonical correlation analysis and application towards online structural damage detection.Journal of Sound and Vibration, 468:115101, 2020

2020
[13]

Review of deep learning: concepts, cnn architectures, challenges, applications, future directions.Journal of big Data, 8(1):53, 2021

Laith Alzubaidi, Jinglan Zhang, Amjad J Humaidi, Ayad Al-Dujaili, Ye Duan, Omran Al-Shamma, José Santa- maría, Mohammed A Fadhel, Muthana Al-Amidie, and Laith Farhan. Review of deep learning: concepts, cnn architectures, challenges, applications, future directions.Journal of big Data, 8(1):53, 2021

2021
[14]

Autonomous agentic ai with policy adaptation for physics-informed spectral learning in structural health monitoring.Advanced Engineering Informatics, 70:104224, 2026

Anshu Sharma and Basuraj Bhowmik. Autonomous agentic ai with policy adaptation for physics-informed spectral learning in structural health monitoring.Advanced Engineering Informatics, 70:104224, 2026

2026
[15]

U-net: Convolutional networks for biomedical image segmentation

Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InInternational Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015

2015
[16]

Semi-Supervised Classification with Graph Convolutional Networks

Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907, 2016

work page internal anchor Pith review Pith/arXiv arXiv 2016
[17]

Inductive representation learning on large graphs.Advances in neural information processing systems, 30, 2017

Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs.Advances in neural information processing systems, 30, 2017

2017
[18]

Using graph neural networks and frequency domain data for automated operational modal analysis of populations of structures.Data-Centric Engineering, 6:e45, 2025

Xudong Jian, Yutong Xia, Gregory Duthé, Kiran Bacsa, Wei Liu, and Eleni Chatzi. Using graph neural networks and frequency domain data for automated operational modal analysis of populations of structures.Data-Centric Engineering, 6:e45, 2025

2025
[19]

Kipf et al.Variational Graph Auto- Encoders

Thomas N Kipf and Max Welling. Variational graph auto-encoders.arXiv preprint arXiv:1611.07308, 2016

work page arXiv 2016
[20]

Estimating disper- sion curves from frequency response functions via vector-fitting.Mechanical Systems and Signal Processing, 140:106597, 2020

Mohammad I Albakri, Vijaya VN Sriram Malladi, Serkan Gugercin, and Pablo A Tarazaga. Estimating disper- sion curves from frequency response functions via vector-fitting.Mechanical Systems and Signal Processing, 140:106597, 2020

2020
[21]

Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning

Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning.(jun 2015).arXiv preprint arXiv:1506.02142, 2015

work page Pith review arXiv 2015
[22]

Simple and scalable predictive uncertainty estimation using deep ensembles

B Lakshminarayanan, A Pritzel, and C Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. arxiv [stat. ml], 2016

2016
[23]

Auto-Encoding Variational Bayes

Diederik P Kingma and Max Welling. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013. 22 Physics-Aware U-Res VGAE

work page internal anchor Pith review Pith/arXiv arXiv 2013
[24]

The modal assurance criterion–twenty years of use and abuse.Sound and vibration, 37(8):14–23, 2003

Randall J Allemang. The modal assurance criterion–twenty years of use and abuse.Sound and vibration, 37(8):14–23, 2003

2003
[25]

A general framework for quantifying aleatoric and epistemic uncertainty in graph neural networks, 2022

Sai Munikoti, Deepesh Agarwal, Laya Das, and Balasubramaniam Natarajan. A general framework for quantifying aleatoric and epistemic uncertainty in graph neural networks, 2022

2022
[26]

Hamilton, Rex Ying, and Jure Leskovec

William L. Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large graphs, 2018

2018
[27]

Resgraphnet: Graphsage with embedded residual module for prediction of global monthly mean temperature.Artificial Intelligence in Geosciences, 3:148–156, 2022

Ziwei Chen, Zhiguo Wang, Yang Yang, and Jinghuai Gao. Resgraphnet: Graphsage with embedded residual module for prediction of global monthly mean temperature.Artificial Intelligence in Geosciences, 3:148–156, 2022

2022
[28]

Network in graph neural network, 2021

Xiang Song, Runjie Ma, Jiahang Li, Muhan Zhang, and David Paul Wipf. Network in graph neural network, 2021

2021
[29]

Itoh, Takatomi Kubo, and Kazushi Ikeda

Takeshi D. Itoh, Takatomi Kubo, and Kazushi Ikeda. Multi-level attention pooling for graph neural networks: Unifying graph representations with multiple localities.Neural Networks, 145:356–373, January 2022

2022
[30]

Variational graph normalized autoencoders

Seong Jin Ahn and MyoungHo Kim. Variational graph normalized autoencoders. InProceedings of the 30th ACM international conference on information & knowledge management, pages 2827–2831, 2021

2021
[31]

Deep evidential regression.Advances in neural information processing systems, 33:14927–14937, 2020

Alexander Amini, Wilko Schwarting, Ava Soleimany, and Daniela Rus. Deep evidential regression.Advances in neural information processing systems, 33:14927–14937, 2020

2020
[32]

Trustworthy multimodal regression with mixture of normal-inverse gamma distributions.Advances in Neural Information Processing Systems, 34:6881–6893, 2021

Huan Ma, Zongbo Han, Changqing Zhang, Huazhu Fu, Joey Tianyi Zhou, and Qinghua Hu. Trustworthy multimodal regression with mixture of normal-inverse gamma distributions.Advances in Neural Information Processing Systems, 34:6881–6893, 2021

2021
[33]

A comprehensive survey on evidential deep learning and its applications.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

Junyu Gao, Mengyuan Chen, Liangyu Xiang, and Changsheng Xu. A comprehensive survey on evidential deep learning and its applications.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

2025
[34]

A simple baseline for bayesian uncertainty in deep learning.Advances in neural information processing systems, 32, 2019

Wesley J Maddox, Pavel Izmailov, Timur Garipov, Dmitry P Vetrov, and Andrew Gordon Wilson. A simple baseline for bayesian uncertainty in deep learning.Advances in neural information processing systems, 32, 2019

2019
[35]

Estimation of the continuous ranked probability score with limited informa- tion and applications to ensemble weather forecasts.Mathematical Geosciences, 50(2):209–234, 2018

Michaël Zamo and Philippe Naveau. Estimation of the continuous ranked probability score with limited informa- tion and applications to ensemble weather forecasts.Mathematical Geosciences, 50(2):209–234, 2018

2018
[36]

On calibration of modern neural networks

Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321–1330. PMLR, 2017

2017
[37]

Peter Welch. The use of fast fourier transform for the estimation of power spectra: A method based on time averaging over short, modified periodograms.IEEE Transactions on audio and electroacoustics, 15(2):70–73, 1967

1967
[38]

Understanding model calibration–a gentle introduction and visual exploration of calibration and the expected calibration error (ece).arXiv preprint arXiv:2501.19047, 2025

Maja Pavlovic. Understanding model calibration–a gentle introduction and visual exploration of calibration and the expected calibration error (ece).arXiv preprint arXiv:2501.19047, 2025. 23

work page arXiv 2025