Article
Engineering, Multidisciplinary
F. Llorente, E. Curbelo, L. Martino, V. Elvira, D. Delgado
Summary: This study focuses on layered adaptive importance sampling algorithms and proposes different enhancements to increase efficiency and reduce computational costs. Strategies for designing cheaper schemes are also introduced. Numerical experiments demonstrate the advantages of the proposed schemes in handling computational challenges in real-world applications.
APPLIED MATHEMATICAL MODELLING
(2022)
Article
Physics, Multidisciplinary
Maximilian Dax, Stephen R. Green, Jonathan Gair, Michael Purrer, Jonas Wildberger, Jakob H. Macke, Alessandra Buonanno, Bernhard Scholkopf
Summary: We combine amortized neural posterior estimation with importance sampling to achieve fast and accurate gravitational-wave inference. This approach addresses criticisms against deep learning for scientific inference by providing a corrected posterior, a performance diagnostic for proposal assessment, and an unbiased estimate of the Bayesian evidence. Our study of 42 binary black hole mergers shows a significant improvement in sample efficiency and reduction in statistical uncertainty, indicating the potential impact of this method in gravitational-wave inference and other scientific applications.
PHYSICAL REVIEW LETTERS
(2023)
Article
Multidisciplinary Sciences
Ben Lambert, Chon Lok Lei, Martin Robinson, Michael Clerx, Richard Creswell, Sanmitra Ghosh, Simon Tavener, David J. Gavaghan
Summary: Ordinary differential equation models are commonly used in biology to describe dynamic processes. However, when performing likelihood-based parameter inference on these models, the contribution of factors not explicitly included in the mathematical model needs to be represented. Independent Gaussian noise is often chosen as a statistical process, assuming random latent factors. This study demonstrates that highly persistent differences between observations and model quantities can occur, and using the wrong noise assumption can lead to artificial reduction of uncertainty in parameter estimation. They propose a workflow to diagnose and remodel for correlated errors.
JOURNAL OF THE ROYAL SOCIETY INTERFACE
(2023)
Review
Multidisciplinary Sciences
Arnaud Doucet, Eric Moulines, Achille Thin
Summary: Latent variable models are popular and have been combined with neural networks to create deep latent variable models. However, the intractability of their likelihood function requires approximations for inference. The article reviews recent strategies such as importance sampling, Markov chain Monte Carlo, and sequential Monte Carlo to improve the bounds of the evidence lower bound (ELBO) for these models.
PHILOSOPHICAL TRANSACTIONS OF THE ROYAL SOCIETY A-MATHEMATICAL PHYSICAL AND ENGINEERING SCIENCES
(2023)
Article
Biology
Jan Boelts, Jan-Matthis Lueckmann, Richard Gao, Jakob H. Macke
Summary: Inferring parameters of computational models is crucial in cognitive neuroscience. Simulation-based inference (SBI) using neural density estimators provides a more efficient way to capture decision-making data. Compared to traditional methods, this approach demonstrates higher accuracy and training efficiency.
Article
Statistics & Probability
Martin Outzen Berild, Sara Martino, Virgilio Gomez-Rubio, Havard Rue
Summary: This article introduces the integrated nested Laplace approximation (INLA) method and its extensions, such as the importance sampling combined with INLA (IS-INLA) and the adaptive multiple importance sampling combined with INLA (AMIS-INLA). The performance of these methods is evaluated through comparison experiments on simulated and observed datasets. The results show that the AMIS-INLA method outperforms other methods in terms of accuracy, efficiency, and robustness, although the IS-INLA algorithm can be considered for faster inference when good proposals are available.
JOURNAL OF COMPUTATIONAL AND GRAPHICAL STATISTICS
(2022)
Article
Computer Science, Interdisciplinary Applications
Samuel Wiqvist, Andrew Golightly, Ashleigh T. McLean, Umberto Picchini
Summary: Stochastic differential equation mixed-effects models (SDEMEMs) are hierarchical models that can handle random variability in time dynamics and between experimental units, and also address measurement error. Fully Bayesian inference is performed using a sampling-based method, with a Gibbs sampler targeting the marginal posterior of parameter values. The methodology is computationally efficient and can deal with a variety of SDEMEMs.
COMPUTATIONAL STATISTICS & DATA ANALYSIS
(2021)
Article
Astronomy & Astrophysics
Geraint Pratten, Cecilio Garcia-Quiros, Marta Colleoni, Antoni Ramos-Buades, Hector Estelles, Maite Mateu-Lucena, Rafel Jaume, Maria Haney, David Keitel, Jonathan E. Thompson, Sascha Husa
Summary: IMRPhenomXPHM is a phenomenological frequency-domain model for gravitational-wave signals from quasicircular precessing binary black holes, incorporating higher order multipoles beyond the dominant quadrupole. The model is an extension of IMRPhenomXHM, using approximate maps and specific choices to enhance accuracy and speed. With reduced computational cost and improved interpolation techniques, it provides a productive tool for gravitational wave astronomy in the current era of increased event detections.
Article
Multidisciplinary Sciences
Yuting Li, Guenther Turk, Paul B. Rohrbach, Patrick Pietzonka, Julian Kappler, Rajesh Singh, Jakub Dolezal, Timothy Ekeh, Lukas Kikuchi, Joseph D. Peterson, Austen Bolitho, Hideki Kobayashi, Michael E. Cates, R. Adhikari, Robert L. Jack
Summary: The study presents a Bayesian inference methodology for quantifying uncertainties in epidemiological forecasts, specifically for epidemics modeled by non-stationary, continuous-time, Markov population processes. The method's efficiency is derived from an approximation of the likelihood using a functional central limit theorem, which is valid for large populations. The methodology is demonstrated by analyzing the early stages of the COVID-19 pandemic in the UK, utilizing age-structured data for deaths.
ROYAL SOCIETY OPEN SCIENCE
(2021)
Article
Engineering, Multidisciplinary
Chaolin Song, Zeyu Wang, Abdollah Shafieezadeh, Rucheng Xiao
Summary: Bayesian updating combined with structural reliability methods provides an efficient and accurate framework for probabilistic calibration. This paper addresses the issue of insufficient sample size in Bayesian updating by introducing an importance sampling density based on Gaussian mixture distribution and an active learning framework. Additionally, it discusses the estimation of the posterior distribution and proposes a stopping criterion.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2022)
Article
Engineering, Civil
Kaixuan Feng, Zhenzhou Lu, Jiaqi Wang, Pengfei He, Ying Dai
Summary: Reliability updating is an effective tool to reassess system reliability when new observation information is obtained. The adaptive Kriging based reliability updating method (RUAK) improves the efficiency of reliability updating by incorporating adaptive Kriging into traditional simulation. However, the use of an identical candidate sampling pool in RUAK for estimating both prior and posterior failure probabilities wastes computational resources when there are significant differences between the importance regions in estimation.
Article
Ecology
Dave W. Hudson, Dave J. Hodgson, Michael A. Cant, Faye J. Thompson, Richard Delahay, Robbie A. McDonald, Trevelyan J. McKinley
Summary: This paper introduces a Bayesian approach to ecological system modeling, focusing on the task of multimodel inference through estimating posterior model weights. The Importance Sampling method used for model comparison aligns with RJ-MCMC model comparisons and is often more straightforward to fit and optimize.
METHODS IN ECOLOGY AND EVOLUTION
(2023)
Article
Agronomy
Pratishtha Poudel, Nora M. Bello, Romulo P. Lollato, Phillip D. Alderman
Summary: The goal of this study was to use a dynamic ordinary differential equation (ODE) model within a Bayesian framework for stochastic inference on system-level parameters, and compare its predictive performance to more commonly used modeling approaches for repeated measures data. The results showed that none of the modeling approaches clearly outperformed any other in terms of goodness of fit or prediction accuracy.
FIELD CROPS RESEARCH
(2022)
Article
Engineering, Electrical & Electronic
F. Llorente, L. Martino, D. Delgado-Gomez, G. Camps-Valls
Summary: This study introduces a AIS framework called RADIS, which minimizes the mismatch between proposal and target densities by adaptively constructing a non-parametric proposal density. RADIS is based on a deep architecture of two (or more) nested IS schemes to draw samples from the constructed emulator. By adding more support points, RADIS asymptotically converges to an exact sampler under mild conditions.
DIGITAL SIGNAL PROCESSING
(2021)
Article
Astronomy & Astrophysics
Marta Colleoni, Maite Mateu-Lucena, Hector Estelles, Cecilio Garcia-Quiros, David Keitel, Geraint Pratten, Antoni Ramos-Buades, Sascha Husa
Summary: In this study, the authors reanalyze the gravitational-wave event GW190412 using state-of-the-art phenomenological waveform models, focusing on the contribution from subdominant harmonics. They compare the PhenomX and PhenomT waveform models, discussing their construction techniques, computational efficiency, and agreement with other waveform models. Additionally, practical aspects of Bayesian inference, such as run convergence and computational cost, are also discussed.