Search NASA⌕ Search

SEARCH · Search NASA

Results for “Bayesian approach”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 289 records · Page 16

Determining the Bayesian optimal sampling strategy in a hierarchical system.

Consider a classic hierarchy tree as a basic model of a 'system-of-systems' network, where each node represents a component system (which may itself consist of a set of sub-systems). For this general composite system, we present a technique for computing the optimal testing strategy, which is based on Bayesian decision analysis. In previous work, we developed a Bayesian approach for computing the distribution of the reliability of a system-of-systems structure that uses test data and prior information. This allows for the determination of both an estimate of the reliability and a quantification of confidence in the estimate. Improving the accuracy of the reliability estimate and increasing the corresponding confidence require the collection of additional data. However, testing all possible sub-systems may not be cost-effective, feasible, or even necessary to achieve an improvement in the reliability estimate. To address this sampling issue, we formulate a Bayesian methodology that systematically determines the optimal sampling strategy under specified constraints and costs that will maximally improve the reliability estimate of the composite system, e.g., by reducing the variance of the reliability distribution. This methodology involves calculating the 'Bayes risk of a decision rule' for each available sampling strategy, where risk quantifies the relative effect that each sampling strategy could have on the reliability estimate. A general numerical algorithm is developed and tested using an example multicomponent system. The results show that the procedure scales linearly with the number of components available for testing.

99 GENERAL AND MISCELLANEOUS↗

Developing an Evolutionary Baseline Model for Humans: Jointly Inferring Purifying Selection with Population History

Building evolutionarily appropriate baseline models for natural populations is not only important for answering fundamental questions in population genetics—including quantifying the relative contributions of adaptive versus nonadaptive processes—but also essential for identifying candidate loci experiencing relatively rare and episodic forms of selection (e.g., positive or balancing selection). Here, a baseline model was developed for a human population of West African ancestry, the Yoruba, comprising processes constantly operating on the genome (i.e., purifying and background selection, population size changes, recombination rate heterogeneity, and gene conversion). Specifically, to perform joint inference of selective effects with demography, an approximate Bayesian approach was employed that utilizes the decay of background selection effects around functional elements, taking into account genomic architecture. This approach inferred a recent 6-fold population growth together with a distribution of fitness effects that is skewed towards effectively neutral mutations. Importantly, these results further suggest that, although strong and/or frequent recurrent positive selection is inconsistent with observed data, weak to moderate positive selection is consistent but unidentifiable if rare.

59 BASIC BIOLOGICAL SCIENCES↗

Forward variable selection enables fast and accurate dynamic system identification with Karhunen-Loève decomposed Gaussian processes

A promising approach for scalable Gaussian processes (GPs) is the Karhunen-Loève (KL) decomposition, in which the GP kernel is represented by a set of basis functions which are the eigenfunctions of the kernel operator. Such decomposed kernels have the potential to be very fast, and do not depend on the selection of a reduced set of inducing points. However KL decompositions lead to high dimensionality, and variable selection thus becomes paramount. This paper reports a new method of forward variable selection, enabled by the ordered nature of the basis functions in the KL expansion of the Bayesian Smoothing Spline ANOVA kernel (BSS-ANOVA), coupled with fast Gibbs sampling in a fully Bayesian approach. It quickly and effectively limits the number of terms, yielding a method with competitive accuracies, training and inference times for tabular datasets of low feature set dimensionality. Theoretical computational complexities are O ( N P 2 ) in training and O ( P ) per point in inference, where N is the number of instances and P the number of expansion terms. The inference speed and accuracy makes the method especially useful for dynamic systems identification, by modeling the dynamics in the tangent space as a static problem, then integrating the learned dynamics using a high-order scheme. The methods are demonstrated on two dynamic datasets: a ‘Susceptible, Infected, Recovered’ (SIR) toy problem, along with the experimental ‘Cascaded Tanks’ benchmark dataset. Comparisons on the static prediction of time derivatives are made with a random forest (RF), a residual neural network (ResNet), and the Orthogonal Additive Kernel (OAK) inducing points scalable GP, while for the timeseries prediction comparisons are made with LSTM and GRU recurrent neural networks (RNNs) along with the SINDy package.

Hayes, Kyle↗

Composite Bond Line Measurements Based on A Bayesian Analysis of Flash Thermography Data

For bonded composite materials, an accurate characterization of the adhesive bond line is needed to predict failure modes and fracture toughness. In this paper, bond line thickness was estimated from data obtained using through transmission flash thermography. The forward model that predicts back surface temperature is based on a three layer heat diffusion equation with varying diffusivity and flux boundary conditions. The corresponding inverse problem of estimating bond line thickness from measurement data was solved using a Bayesian approach that assumed Gaussian priors for the bond line thickness and thermal diffusivity of the adherends. Finally, the outputs of the thermography based method were compared to measurements that were collected using a micrometer and ultrasound testing.

Flash thermography↗

Exact and efficient hybrid Monte Carlo algorithm for accelerated Bayesian inference of gene expression models from snapshots of single-cell transcripts

Single cells exhibit a significant amount of variability in transcript levels, which arises from slow, stochastic transitions between gene expression states. Elucidating the nature of these states and understanding how transition rates are affected by different regulatory mechanisms require state-of-the-art methods to infer underlying models of gene expression from single cell data. A Bayesian approach to statistical inference is the most suitable method for model selection and uncertainty quantification of kinetic parameters using small data sets. However, this approach is impractical because current algorithms are too slow to handle typical models of gene expression. To solve this problem, we first show that time-dependent mRNA distributions of discrete-state models of gene expression are dynamic Poisson mixtures, whose mixing kernels are characterized by a piecewise deterministic Markov process. Here, we combined this analytical result with a kinetic Monte Carlo algorithm to create a hybrid numerical method that accelerates the calculation of time-dependent mRNA distributions by 1000-fold compared to current methods. We then integrated the hybrid algorithm into an existing Monte Carlo sampler to estimate the Bayesian posterior distribution of many different, competing models in a reasonable amount of time. We demonstrate that kinetic parameters can be reasonably constrained for modestly sampled data sets if the model is known a priori. If there are many competing models, Bayesian evidence can rigorously quantify the likelihood of a model relative to other models from the data. We demonstrate that Bayesian evidence selects the true model and outperforms approximate metrics typically used for model selection.

59 BASIC BIOLOGICAL SCIENCES↗

Measurement of the electroweak top quark production cross section and the CKM matrix element Vtb with the D0 experiment

At particle accelerators the Standard Model has been tested and will be tested further to a great precision. The data analyzed in this thesis have been collected at the world's highest energetic-collider, the Tevatron, located at the Fermi National Accelerator Laboratory (FNAL) in the vicinity of Chicago, IL, USA. There, protons and antiprotons are collided at a center-of-mass energy of {radical}s = 1.96 TeV. The discovery of the top quark was one of the remarkable results not only for the CDF and D0 experiments at the Tevatron collider, but also for the Standard Model, which had predicted the existence of the top quark because of symmetry arguments long before already. Still, the Tevatron is the only facility able to produce top quarks. The predominant production mechanism of top quarks is the production of a top-antitop quark pair via the strong force. However, the Standard Model also allows the production of single top quarks via the electroweak interaction. This process features the unique opportunity to measure the |V tb | matrix element of the Cabbibo-Kobayashi-Maskawa (CKM) matrix directly, without assuming unitarity of the matrix or assuming that the number of quark generations is three. Hence, the measurement of the cross section of electroweak top quark production is more than the technical challenge to extract a physics process that only occurs one out of ten billion collisions. It is also an important test of the V-A structure of the electroweak interaction and a potential window to physics beyond the Standard Model in the case where the measurement of |V{sub tb}| would result in a value significantly different from 1, the value predicted by the Standard Model. At the Tevatron two production processes contribute significantly to the production of single top quarks: the production via the t-channel, also called W-gluon fusion, and the production via the s-channel, known as well as W* process. This analysis searches for the combined s+t channel production cross section, assuming the ratio of s-channel production over t-channel production is realized in nature as predicted by the Standard Model. A data set of approximately 1 fb -1 is analyzed, the data set used by the D0 collaboration to claim evidence for single top quark production. Events with two, three, and four jets are used in the analysis if they contain one or two jets that were tagged as originating from the decay of a b hadron, an isolated muon or electron, and a significant amount of missing transverse energy. This selection of events follows the signature that the single top quark events are expected to show in the detector. In the meantime, both collaborations D0 and CDF have analyzed a larger data set and have celebrated the joint observation of single top quark production. The novelty of the analysis presented here is the way discriminating observables are determined. A so-called Multi-Process Factory evaluates each event under several hypotheses. A common analysis technique for example in top quark properties studies is to reconstruct the intermediate particles in the decay chain of the signal process from the final state objects measured in the various subdetectors. An essential part of such a method is to resolve the ambiguities that arise in the assignment of the final state objects to the partons of the decay chain. In a Multi-Process Factory this approach is extended and not only the decay chain of the signal process is reconstructed, but also the decay chains of the most important background processes. From the numerous possible event configurations for each of the signal and background decay chains the most probable configuration is selected based on a likelihood measure. Properties of this configuration, such as mass of the reconstructed top quark, are then used in a multivariate analysis technique to separate the expected signal contribution from the background processes. The technique which is used is called Boosted Decision Trees and has only recently been introduced in high energy physics analyses. A Bayesian approach is used to finally extract the cross section from the discriminant output of Boosted Decision Trees.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

Bayesian Estimation of Earth’s Undiscovered Mineralogical Diversity Using Noninformative Priors

Recently, statistical distributions have been explored to provide estimates of the mineralogical diversity of Earth, and Earth-like planets. In this paper, a Bayesian approach is introduced to estimate Earth’s undiscovered mineralogical diversity. Samples are generated from a posterior distribution of the model parameters using Markov chain Monte Carlo simulations such that estimates and inference are directly obtained. It was previously shown that the mineral species frequency distribution conforms to a generalized inverse Gauss–Poisson (GIGP) large number of rare events model. Even though the model fit was good, the population size estimate obtained by using this model was found to be unreasonably low by mineralogists. In this paper, several zero-truncated, mixed Poisson distributions are fitted and compared, where the Poisson-lognormal distribution is found to provide the best fit. Subsequently, the population size estimates obtained by Bayesian methods are compared to the empirical Bayes estimates. Species accumulation curves are constructed and employed to estimate the population size as a function of sampling size. Finally, the relative abundances, and hence the occurrence probabilities of species in a random sample, are calculated numerically for all mineral species in Earth’s crust using the Poisson-lognormal distribution. These calculations are connected and compared to the calculations obtained in a previous paper using the GIGP model for which mineralogical criteria of an Earth-like planet were given.

Bayesian statistics↗

Optimization of an aerostructural machining process using physics-guided Bayesian stability modelling

Existing algorithms for predicting milling chatter have not been widely adopted in industry since they require specialized instruments to measure the stability inputs. This study describes how the machining process for a meter-scale aluminum aerostructure was optimized using a physics-guided Bayesian stability model. The study was performed in collaboration with an industrial partner on production machines to evaluate the practicality of the proposed method under real-world conditions. For each cutting tool, the Bayesian approach automatically selected a small number of cutting tests, which were monitored using a microphone to observe the chatter frequency. The algorithm learned the system dynamics, cutting forces, and stability map from these test results. A novel algorithm for predicting tool bending stress was incorporated into the test selection algorithm to avoid tool breakage. On average, each set of optimized cutting parameters required less than six tests to identify and were 97% more productive than baseline parameters from the cutting tool manufacturer. The machining program was then further optimized using commercial feedrate scheduling software to remove cutting force spikes and reduce air cutting time. Five components were machined using the optimized process. These results demonstrate the potential for physics-guided Bayesian models to improve productivity in industrial settings.

Cornelius, Aaron [UT Knoxville]↗

Goal-oriented a-posteriori estimation of model error as an aid to parameter estimation

In this work, a Bayesian model calibration framework is presented that utilizes goal-oriented a-posterior error estimates in quantities of interest (QoIs) for classes of high-fidelity models characterized by PDEs. It is shown that for a large class of computational models, it is possible to develop a computationally inexpensive procedure for calibrating parameters of high-fidelity models of physical events when the parameters of low-fidelity (surrogate) models are known with acceptable accuracy. The main ingredients in the proposed model calibration scheme are goal-oriented a-posteriori estimates of error in QoIs computed using a so-called lower fidelity model compared to those of an uncalibrated higher fidelity model. The estimates of error in QoIs are used to define likelihood functions in Bayesian inversion analysis. A standard Bayesian approach is employed to compute the posterior distribution of model parameters of high-fidelity models. As applications, parameters in a quasi-linear second-order elliptic boundary-value problem (BVP) are calibrated using a second-order linear elliptic BVP. In a second application, parameters of a tumor growth model involving nonlinear time-dependent PDEs are calibrated using a lower fidelity linear tumor growth model with known parameter values.

A-posterior estimates↗

Bayesian Optimization of a Free-Electron Laser

The Linac Coherent Light Source X-ray free-electron laser is a complex scienti c apparatus which changes con gurations multiple times per day, necessitating fast tuning strategies to reduce setup time for successive experiments. To this end, we employ a Bayesian approach to maximizing X- ray laser pulse energy by controlling groups of quadrupole magnets. A Gaussian process model provides probabilistic predictions for the machine response with respect to control parameters, enabling a balance of exploration and exploitation in the search for the global optimum. We show that the model parameters can be learned from archived scans, and correlations between devices can be extracted from the beam transport. Furthermore, the result is a sample-e cient optimization routine, combining both historical data and knowledge of accelerator physics to signi cantly outperform existing optimizers.

47 OTHER INSTRUMENTATION↗

WE-E-BRE-05: Ensemble of Graphical Models for Predicting Radiation Pneumontis Risk

Purpose: We propose a prior knowledge-based approach to construct an interaction graph of biological and dosimetric radiation pneumontis (RP) covariates for the purpose of developing a RP risk classifier. Methods: We recruited 59 NSCLC patients who received curative radiotherapy with minimum 6 month follow-up. 16 RP events was observed (CTCAE grade ≥2). Blood serum was collected from every patient before (pre-RT) and during RT (mid-RT). From each sample the concentration of the following five candidate biomarkers were taken as covariates: alpha-2-macroglobulin (α2M), angiotensin converting enzyme (ACE), transforming growth factor β (TGF-β), interleukin-6 (IL-6), and osteopontin (OPN). Dose-volumetric parameters were also included as covariates. The number of biological and dosimetric covariates was reduced by a variable selection scheme implemented by L1-regularized logistic regression (LASSO). Posterior probability distribution of interaction graphs between the selected variables was estimated from the data under the literature-based prior knowledge to weight more heavily the graphs that contain the expected associations. A graph ensemble was formed by averaging the most probable graphs weighted by their posterior, creating a Bayesian Network (BN)-based RP risk classifier. Results: The LASSO selected the following 7 RP covariates: (1) pre-RT concentration level of α2M, (2) α2M level mid- RT/pre-RT, (3) pre-RT IL6 level, (4) IL6 level mid-RT/pre-RT, (5) ACE mid-RT/pre-RT, (6) PTV volume, and (7) mean lung dose (MLD). The ensemble BN model achieved the maximum sensitivity/specificity of 81%/84% and outperformed univariate dosimetric predictors as shown by larger AUC values (0.78∼0.81) compared with MLD (0.61), V20 (0.65) and V30 (0.70). The ensembles obtained by incorporating the prior knowledge improved classification performance for the ensemble size 5∼50. Conclusion: We demonstrated a probabilistic ensemble method to detect robust associations between RP covariates and its potential to improve RP prediction accuracy. Our Bayesian approach to incorporate prior knowledge can enhance efficiency in searching of such associations from data. The authors acknowledge partial support by: 1) CREATE Medical Physics Research Training Network grant of the Natural Sciences and Engineering Research Council (Grant number: 432290) and 2) The Terry Fox Foundation Strategic Training Initiative for Excellence in Radiation Research for the 21st Century (EIRR21)

60 APPLIED LIFE SCIENCES↗

A model-independent analysis of the Fermi Large Area Telescope gamma-ray data from the Milky Way dwarf galaxies and halo to constrain dark matter scenarios

Here, we implemented a novel technique to perform the collective spectral analysis of sets of multiple gamma-ray point sources using the data collected by the Large Area Telescope onboard the Fermi satellite. The energy spectra of the sources are reconstructed starting from the photon counts and without assuming any spectral model for both the sources and the background. In case of faint sources, upper limits on their fluxes are evaluated with a Bayesian approach. Our analysis technique is very useful when several sources with similar spectral features are studied, such as sources of gamma rays from annihilation of dark matter particles. We also present the results obtained by applying this analysis to a sample of dwarf spheroidal galaxies and to the Milky Way dark matter halo. The analysis of dwarf spheroidal galaxies yields upper limits on the product of the dark matter pair annihilation cross section and the relative velocity of annihilating particles that are well below those predicted by the canonical thermal relic scenario in a mass range from a few GeV to a few tens of GeV for some annihilation channels.

79 ASTRONOMY AND ASTROPHYSICS↗

Bayesian Calibration of Nuclear Graphite Property Models Accounting for Model Inadequacy and Impacts on Component Performance

Nuclear-grade structural graphite is extensively utilized in the core designs of various advanced nuclear reactors. In the reactor environment, graphite is subjected to prolonged exposure to extreme conditions, including high temperatures, radiation, and potentially molten salt and oxygen. Such exposure can induce several degradation mechanisms in graphite, including nonuniform volumetric strains caused by irradiation and thermal expansion, leading to stresses that may compromise the performance of graphite components. Assessing component integrity requires accurate models of graphite's thermomechanical response. This report documents the Bayesian calibration of thermomechanical properties for nuclear-grade graphite and their application to graphite component modeling and simulation using the Grizzly code. As part of this work, uncertainty-quantified models were developed for the elastic modulus, coefficient of thermal expansion, irradiation-induced dimensional change, and irradiation-induced creep for graphite grades IG-110, NBG-18, NBG-17, PCEA, and 2114. Using a hierarchical Bayesian approach, multiple experimental data sources were combined to develop Gaussian process models for the properties. Using the Kennedy O'Hagan framework, the uncertainties due to inadequacies in the model and the inherent spread in the experimental data were quantified for three different models. These uncertainty-quantified models, with a model-form correction, were subsequently applied to a coupled-physics simulation of representative graphite components, revealing that the uncertainties have a large impact on the components' deformation.

36 - MATERIALS SCIENCE↗

Multimodal parameter spaces of a complex multi-channel neuron model

One of the most common types of models that helps us to understand neuron behavior is based on the Hodgkin–Huxley ion channel formulation (HH model). A major challenge with inferring parameters in HH models is non-uniqueness: many different sets of ion channel parameter values produce similar outputs for the same input stimulus. Such phenomena result in an objective function that exhibits multiple modes (i.e., multiple local minima). This non-uniqueness of local optimality poses challenges for parameter estimation with many algorithmic optimization techniques. HH models additionally have severe non-linearities resulting in further challenges for inferring parameters in an algorithmic fashion. To address these challenges with a tractable method in high-dimensional parameter spaces, we propose using a particular Markov chain Monte Carlo (MCMC) algorithm, which has the advantage of inferring parameters in a Bayesian framework. The Bayesian approach is designed to be suitable for multimodal solutions to inverse problems. We introduce and demonstrate the method using a three-channel HH model. We then focus on the inference of nine parameters in an eight-channel HH model, which we analyze in detail. We explore how the MCMC algorithm can uncover complex relationships between inferred parameters using five injected current levels. The MCMC method provides as a result a nine-dimensional posterior distribution, which we analyze visually with solution maps or landscapes of the possible parameter sets. The visualized solution maps show new complex structures of the multimodal posteriors, and they allow for selection of locally and globally optimal value sets, and they visually expose parameter sensitivities and regions of higher model robustness. We envision these solution maps as enabling experimentalists to improve the design of future experiments, increase scientific productivity and improve on model structure and ideation when the MCMC algorithm is applied to experimental data.

97 MATHEMATICS AND COMPUTING↗

Investigation of Ethane Dehydrogenation and Hydrogenolysis on Pt(111), Pt(211), and Pt(100): Bayesian Quantification and Correction of DFT-Based Enthalpic and Entropic Uncertainties

Computational investigations of heterogeneously catalyzed reactions using density functional theory (DFT) are often inaccurate, largely due to uncertainties in the choice of DFT functional (enthalpic uncertainty) and approximations for modeling adsorbate movement along the catalyst surface (entropic uncertainty). This work illustrates that both uncertainties are significant in the investigation of ethane dehydrogenation (EDH) and hydrogenolysis on Pt catalysts by considering the complete deconstruction of ethane on Pt(111), Pt(211), and Pt(100) using microkinetic modeling (MKM). Hence, this work uses both noncalibrated and Bayesian-calibrated MKMs to quantify and correct inaccuracies in macroscopic properties due to both uncertainties. A Bayesian approach to the correction of entropic errors was introduced using a “Modified Fermi Function (MFF)” to calibrate between the two bounds of entropy represented by the harmonic oscillator (HO) and free translator (FT) approximations. Regardless of enthalpic and entropic uncertainties, all three surfaces are capable of ethane activation; however, Pt(211) was found to be the most active and is largely responsible for methane production. Next, Pt(111) is largely responsible for acetylene production, and Pt(100) has the highest ethylene selectivity but is most susceptible to coking. By comparison of different calibrated models, the FT entropy approximation was found to better describe EDH under typical experimental conditions. Statistical evidence was found to support Pt(111) as the active site for EDH, assuming that one single site is responsible for the chemistry. On the three surfaces, competing second dehydrogenations to CH 2 CH 2 and CH 3 CH were observed as well as isomerization of CH 3 CH back to CH 2 CH 2 and deeper dehydrogenation of CH 3 CH. In conclusion, C–C cleavage was found to largely proceed via the CH 3 C intermediate on Pt(100) and Pt(111), while on Pt(211), it was via both CHC and CH 3 C.

Bayesian model selection↗

An assessment of Earth's climate sensitivity using multiple lines of evidence

We assess evidence relevant to Earth's equilibrium climate sensitivity per doubling of atmospheric CO2, characterized by an effective sensitivity S . This evidence includes feedback process understanding, the historical climate record, and the paleoclimate record. An S value lower than 2 K is difficult to reconcile with any of the three lines of evidence. The amount of cooling during the Last Glacial Maximum provides strong evidence against values of S greater than 4.5 K. Other lines of evidence in combination also show that this is relatively unlikely. We use a Bayesian approach to produce a probability density (PDF) for S given all the evidence, including tests of robustness to difficult‐to‐quantify uncertainties and different priors. The 66% range is 2.6‐3.9 K for our Baseline calculation, and remains within 2.3‐4.5 K under the robustness tests; corresponding 5‐95% ranges are 2.3‐4.7 K, bounded by 2.0‐5.7 K (although such high‐confidence ranges should be regarded more cautiously). This indicates a stronger constraint on S than reported in past assessments, by lifting the low end of the range. This narrowing occurs because the three lines of evidence agree and are judged to be largely independent, and because of greater confidence in understanding feedback processes and in combining evidence. We identify promising avenues for further narrowing the range in S , in particular using comprehensive models and process understanding to address limitations in the traditional forcing‐feedback paradigm for interpreting past changes.

Climate↗

An assessment of Earth's climate sensitivity using multiple lines of evidence

In this work, we assess evidence relevant to Earth's equilibrium climate sensitivity per doubling of atmospheric CO 2 , characterized by an effective sensitivity S . This evidence includes feedback process understanding, the historical climate record, and the paleoclimate record. An S value lower than 2 K is difficult to reconcile with any of the three lines of evidence. The amount of cooling during the Last Glacial Maximum provides strong evidence against values of S greater than 4.5 K. Other lines of evidence in combination also show that this is relatively unlikely. We use a Bayesian approach to produce a probability density (PDF) for S given all the evidence, including tests of robustness to difficult-to-quantify uncertainties and different priors. The 66% range is 2.6-3.9 K for our Baseline calculation, and remains within 2.3-4.5 K under the robustness tests; corresponding 5-95% ranges are 2.3-4.7 K, bounded by 2.0-5.7 K (although such high-confidence ranges should be regarded more cautiously). This indicates a stronger constraint on S than reported in past assessments, by lifting the low end of the range. This narrowing occurs because the three lines of evidence agree and are judged to be largely independent, and because of greater confidence in understanding feedback processes and in combining evidence. We identify promising avenues for further narrowing the range in S , in particular using comprehensive models and process understanding to address limitations in the traditional forcing-feedback paradigm for interpreting past changes.

54 ENVIRONMENTAL SCIENCES↗

An adaptive Hessian approximated stochastic gradient MCMC method

Bayesian approaches have been successfully integrated into training deep neural networks. One popular family is stochastic gradient Markov chain Monte Carlo methods (SG-MCMC), which have gained increasing interest due to their ability to handle large datasets and the potential to avoid overfitting. Although standard SG-MCMC methods have shown great performance in a variety of problems, they may be inefficient when the random variables in the target posterior densities have scale differences or are highly correlated. Here, we present an adaptive Hessian approximated stochastic gradient MCMC method to incorporate local geometric information while sampling from the posterior. The idea is to apply stochastic approximation (SA) to sequentially update a preconditioning matrix at each iteration. The preconditioner possesses second-order information and can guide the random walk of a sampler efficiently. Instead of computing and saving the full Hessian of the log posterior, we use limited memory of the samples and their stochastic gradients to approximate the inverse Hessian-vector multiplication in the updating formula. Moreover, by smoothly optimizing the preconditioning matrix via SA, our proposed algorithm can asymptotically converge to the target distribution with a controllable bias under mild conditions. To reduce the training and testing computational burden, we adopt a magnitude-based weight pruning method to enforce the sparsity of the network. Our method is user-friendly and demonstrates better learning results compared to standard SG-MCMC updating rules. The approximation of inverse Hessian alleviates storage and computational complexities for large dimensional models. Numerical experiments are performed on several problems, including sampling from 2D correlated distribution, synthetic regression problems, and learning the numerical solutions of heterogeneous elliptic PDE. The numerical results demonstrate great improvement in both the convergence rate and accuracy.

97 MATHEMATICS AND COMPUTING↗