Bayesian Active Learning Research Articles

Construction of Al–Si interatomic potential based on Bayesian active learning

Nanoscale simulations for optimizing the performance and processing of Al–Si alloys are currently facing two major obstacles: the scarcity of high-quality semi-empirical potentials tailored to complex alloy systems, and the prohibitively high computational cost associated with ab initio molecular dynamics simulations. In order to enhance simulation efficiency and accuracy of the Al–Si alloys’ microstructural evolution, this study employs a dynamic active learning technique, FLARE, to develop a non-parametric machine learning potential that combines the high accuracy of density functional theory (DFT) with the efficiency of classical molecular dynamics (MD). Without relying on extensive initial ab initio molecular dynamics data or existing databases, collection of the necessary data is progressively made during the active learning process, thereby constructing a potential capable of accurately simulating the structure and dynamics of high-temperature Al–Si alloys. By comparing with experimental measurements and ab initio molecular dynamics calculations, the high accuracy and computational efficiency of this potential is demonstrated in predicting energy, force, structure, and dynamic properties. The results provide novel theoretical insights for optimizing Al–Si alloy processing and underscore the usefulness of active learning methods in constructing high-accuracy potentials.

Surrogate-assisted push and pull search for expensive constrained multi-objective optimization problems

In many real-world engineering optimizations, a large number of objective and constraint function values often need to be obtained through simulation software or physical experiments, which incurs significant computational costs and/or time expenses. These problems are known as expensive constraint multi-objective optimization problems (ECMOPs). This paper combines the push and pull search (PPS) framework and proposes a surrogate-assisted evolutionary algorithm to solve ECMOPs through Bayesian active learning, naming it the surrogate-assisted PPS (SA-PPS). Specifically, during the push search stage, candidate solutions are selected based on two indicators: hypervolume improvement and objective uncertainty. These aim to quickly guide the population towards the unconstrained Pareto front while ensuring diversity. During the pull search stage, the population is partitioned into many subregions through reference vectors, and different selection strategies are assigned to each subregion based on its state, aiming to guide the population towards the constrained Pareto front while ensuring diversity. Furthermore, we introduce a batch data selection strategy that utilizes Bayesian active learning to enable the surrogate model to focus on regions of interest in the pull search stage. Extensive experimental results have shown that the proposed SA-PPS algorithm exhibits superior convergence and diversity compared to 9 state-of-the-art algorithms across a variety of benchmark problems and a real-world optimization problem.

Active Reinforcement Learning for the Semantic Segmentation of Urban Images

Image segmentation using supervised learning algorithms usually requires large amounts of annotated training data, while urban datasets frequently contain unbalanced classes leading to poor detection of under-represented classes. We investigate the use of a reinforced active learning method to address the limitations of semantic segmentation on complex urban scenes. In this method, an agent learns to select small informative regions of the image to be labeled from a pool of unlabeled data. The agent is represented by a deep Q-Network, where a Markov Decision Process (MDP) is used to formulate the Active Learning problem. We introduced the Frequency Weighted Average IoU (FWA IoU) as the image region selection performance metric to reduce the amount of training data while achieving competitive results. Using the Cityscapes and GTAv urban datasets, three baseline image segmentation networks (FPN, DeepLabV3, DeepLabV3+) trained with image regions selected by the proposed FWA IoU metric performed better compared to baseline region selection by active learning methods such as the Random selection, Entropy-based selection, and Bayesian Active Learning by Disagreement. Training performance equivalent to 98% of the fully supervised segmentation network was achieved by labeling only 8% of the dataset.

Realization of superconducting-magnetic energy storage supported DSTATCOM using deep Bayesian Active Learning

Realization of superconducting-magnetic energy storage supported DSTATCOM using deep Bayesian Active Learning

Toward parametric Bayesian adaptive procedures for multi-frequency categorical loudness scaling.

A series of Bayesian adaptive procedures to estimate loudness growth across a wide frequency range from individual listeners was developed, and these procedures were compared. Simulation experiments were conducted based on multinomial psychometric functions for categorical loudness scaling across ten test frequencies estimated from 61 listeners with normal hearing and 87 listeners with sensorineural hearing loss. Adaptive procedures that optimized the stimulus selection based on the interim estimates of two types of category-boundary models were tested. The first type of model was a phenomenological model of category boundaries adopted from previous research studies, while the other type was a data-driven model derived from a previously collected set of categorical loudness scaling data. An adaptive procedure without Bayesian active learning was also implemented. Results showed that all adaptive procedures provided convergent estimates of the loudness category boundaries and equal-loudness contours between 250 and 8000 Hz. Performing post hoc model fitting, using the data-driven model, on the collected data led to satisfactory accuracies, such that all adaptive procedures tested in the current study, independent of modeling approach and stimulus-selection rules, were able to provide estimates of the equal-loudness-level contours between 20 and 100 phons with root-mean-square errors typically under 6 dB after 100 trials.

Multi-objective Bayesian active learning for MeV-ultrafast electron diffraction

Ultrafast electron diffraction using MeV energy beams(MeV-UED) has enabled unprecedented scientific opportunities in the study of ultrafast structural dynamics in a variety of gas, liquid and solid state systems. Broad scientific applications usually pose different requirements for electron probe properties. Due to the complex, nonlinear and correlated nature of accelerator systems, electron beam property optimization is a time-taking process and often relies on extensive hand-tuning by experienced human operators. Algorithm based efficient online tuning strategies are highly desired. Here, we demonstrate multi-objective Bayesian active learning for speeding up online beam tuning at the SLAC MeV-UED facility. The multi-objective Bayesian optimization algorithm was used for efficiently searching the parameter space and mapping out the Pareto Fronts which give the trade-offs between key beam properties. Such scheme enables an unprecedented overview of the global behavior of the experimental system and takes a significantly smaller number of measurements compared with traditional methods such as a grid scan. This methodology can be applied in other experimental scenarios that require simultaneously optimizing multiple objectives by explorations in high dimensional, nonlinear and correlated systems.

A Bayesian active learning approach to comparative judgement within education assessment

Assessment is a crucial part of education. Traditional marking is a source of inconsistencies and unconscious bias, placing a high cognitive load on the assessors. One approach to address these issues is comparative judgement (CJ). In CJ, the assessor is presented with a pair of items of work, and asked to select the better one. Following a series of comparisons, a rank for any item may be derived using a ranking model, for example, the Bradley-Terry model, based on the pairwise comparisons. While CJ is considered to be a reliable method for conducting marking, there are concerns surrounding its transparency, and the ideal number of pairwise comparisons to generate a reliable estimation of the rank order is not known. Additionally, there have been attempts to generate a method of selecting pairs that should be compared next in an informative manner, but some existing methods are known to have created their own bias within results inflating the reliability metric used within the process. As a consequence, a random selection approach is usually deployed.In this paper, we propose a novel Bayesian approach to CJ (which we call BCJ) for determining the ranks of a range of items under scrutiny alongside a new way to select the pairs to present to the marker(s) using active learning, addressing the key shortcomings of traditional CJ. Furthermore, we demonstrate how the entire approach may provide transparency by providing the user insights into how it is making its decisions and, at the same time, being more efficient. Results from our synthetic experiments confirm that the proposed BCJ combined with entropy-driven active learning pair-selection method is superior (i.e. always equal to or significantly better) than other alternatives, for example, the traditional CJ method with differing selection methods such as uniformly random, or the popular no repeating pairs where pairs are selected in a round-robin fashion. We also find that the more comparisons that are conducted, the more accurate BCJ becomes, which solves the issue the current method has of the model deteriorating if too many comparisons are performed. As our approach can generate the complete predicted rank distribution for an item, we also show how this can be utilised in probabilistically devising a predicted grade, guided by the choice of the assessor. Finally, we demonstrate our approach on a real dataset on assessing GCSE (UK school-level) essays, highlighting the advantages of BCJ over CJ.

Scalar bounded-from-below conditions from Bayesian active learning

We present a procedure leveraging Bayesian deep active learning to rapidly produce highly accurate approximate bounded-from-below conditions for arbitrary renormalizable scalar potentials, in the form of a neural network which may be saved and exported for use in arbitrary parameter space scans. We explore the performance of our procedure on three different scalar potentials with either highly nontrivial or unknown symbolic bounded-from-below conditions (the most general two-Higgs doublet model, the three-Higgs doublet model, and a version of the Georgi-Machacek model without custodial symmetry). We find that we can produce fast and highly accurate binary classifiers for all three potentials. Furthermore, for the potentials for which no known symbolic necessary and sufficient conditions on boundedness-from-below exist, our classifiers substantially outperform some common approximate analytical methods, such as producing tractable sufficient but not necessary conditions or evaluating boundedness-from-below conditions for scenarios in which only a subset of the theory’s fields achieve vacuum expectation values. Our methodology can be readily adapted to any renormalizable scalar field theory. For the community’s use, we have developed a package, BFBrain, which allows for the rapid implementation of our analysis procedure on user-specified scalar potentials with a high degree of customizability. Published by the American Physical Society 2024

Structural reliability analysis with extremely small failure probabilities: A quasi-Bayesian active learning method

The concept of Bayesian active learning has recently been introduced from machine learning to structural reliability analysis. Although several specific methods have been successfully developed, significant efforts are still needed to fully exploit their potential and to address existing challenges. This work proposes a quasi-Bayesian active learning method, called ‘Quasi-Bayesian Active Learning Cubature’, for structural reliability analysis with extremely small failure probabilities. The method is established based on a cleaver use of the Bayesian failure probability inference framework. To reduce the computational burden associated with the exact posterior variance of the failure probability, we propose a quasi posterior variance instead. Then, two critical elements for Bayesian active learning, namely the stopping criterion and the learning function, are developed subsequently. The stopping criterion is defined based on the quasi posterior coefficient of variation of the failure probability, whose numerical solution scheme is also tailored. The learning function is extracted from the quasi posterior variance, with the introduction of an additional parameter that allows multi-point selection and hence parallel distributed processing. By testing on four numerical examples, it is empirically shown that the proposed method can assess extremely small failure probabilities with desired accuracy and efficiency.

Bayesian active learning line sampling with log-normal process for rare-event probability estimation

Line sampling (LS) stands as a powerful stochastic simulation method for structural reliability analysis, especially for assessing small failure probabilities. To further improve the performance of traditional LS, a Bayesian active learning idea has recently been pursued. This work presents another Bayesian active learning alternative, called ‘Bayesian active learning line sampling with log-normal process’ (BAL-LS-LP), to traditional LS. In this method, we assign an LP prior instead of a Gaussian process prior over the distance function so as to account for its non-negativity constraint. Besides, the approximation error between the logarithmic approximate distance function and the logarithmic true distance function is assumed to follow a zero-mean normal distribution. The approximate posterior mean and variance of the failure probability are derived accordingly. Based on the posterior statistics of the failure probability, a learning function and a stopping criterion are developed to enable Bayesian active learning. In the numerical implementation of the proposed BAL-LS-LP method, the important direction can be updated on the fly without re-evaluating the distance function. Four numerical examples are studied to demonstrate the proposed method. Numerical results show that the proposed method can estimate extremely small failure probabilities with desired efficiency and accuracy.

Semi-Bayesian active learning quadrature for estimating extremely low failure probabilities

The Bayesian failure probability inference (BFPI) framework provides a sound basis for developing new Bayesian active learning reliability analysis methods. However, it is still computationally challenging to make use of the posterior variance of the failure probability. This study presents a novel method called ‘semi-Bayesian active learning quadrature’ (SBALQ) for estimating extremely low failure probabilities, which builds upon the BFPI framework. The key idea lies in only leveraging the posterior mean of the failure probability to design two crucial components for active learning — the stopping criterion and learning function. In this context, a new stopping criterion is introduced through exploring the structure of the posterior mean. Besides, we also develop a numerical integration technique named ‘hyper-shell simulation’ to estimate the analytically intractable integrals inherent in the stopping criterion. Furthermore, a new learning function is derived from the stopping criterion and by maximizing it a single point can be identified in each iteration of the active learning phase. To enable multi-point selection and facilitate parallel computing, the proposed learning function is modified by incorporating an influence function. Through five numerical examples, it is demonstrated that the proposed method can assess extremely small failure probabilities with desired efficiency and accuracy.

Bayesian active learning with model selection for spectral experiments

Active learning is a common approach to improve the efficiency of spectral experiments. Model selection from the candidates and parameter estimation are often required in the analysis of spectral experiments. Therefore, we proposed an active learning with model selection method using multiple parametric models as learning models. Important points for model selection and its parameter estimation were actively measured using Bayesian posterior distribution. The present study demonstrated the effectiveness of our proposed method for spectral deconvolution and Hamiltonian selection in X-ray photoelectron spectroscopy.

Partially Bayesian active learning cubature for structural reliability analysis with extremely small failure probabilities

The Bayesian failure probability inference (BFPI) framework provides a well-established Bayesian approach to quantifying our epistemic uncertainty about the failure probability resulting from a limited number of performance function evaluations. However, it is still challenging to perform Bayesian active learning of the failure probability by taking advantage of the BFPI framework. In this work, three Bayesian active learning methods are proposed under the name ‘partially Bayesian active learning cubature’ (PBALC), based on a cleaver use of the BFPI framework for structural reliability analysis, especially when small failure probabilities are involved. Since the posterior variance of the failure probability is computationally expensive to evaluate, the underlying idea is to exploit only the posterior mean of the failure probability to design two critical components for Bayesian active learning, i.e., the stopping criterion and the learning function. On this basis, three sets of stopping criteria and learning functions are proposed, resulting in the three proposed methods PBALC1, PBALC2 and PBALC3. Furthermore, the analytically intractable integrals involved in the stopping criteria are properly addressed from a numerical point of view. Five numerical examples are studied to demonstrate the performance of the three proposed methods. It is found empirically that the proposed methods can assess very small failure probabilities and significantly outperform several existing methods in terms of accuracy and efficiency.

Hybrid representation-enhanced sampling for Bayesian active learning in musculoskeletal segmentation of lower extremities.

Manual annotations for training deep learning models in auto-segmentation are time-intensive. This study introduces a hybrid representation-enhanced sampling strategy that integrates both density and diversity criteria within an uncertainty-based Bayesian active learning (BAL) framework to reduce annotation efforts by selecting the most informative training samples. The experiments are performed on two lower extremity datasets of MRI and CT images, focusing on the segmentation of the femur, pelvis, sacrum, quadriceps femoris, hamstrings, adductors, sartorius, and iliopsoas, utilizing a U-net-based BAL framework. Our method selects uncertain samples with high density and diversity for manual revision, optimizing for maximal similarity to unlabeled instances and minimal similarity to existing training data. We assess the accuracy and efficiency using dice and a proposed metric called reduced annotation cost (RAC), respectively. We further evaluate the impact of various acquisition rules on BAL performance and design an ablation study for effectiveness estimation. In MRI and CT datasets, our method was superior or comparable to existing ones, achieving a 0.8% dice and 1.0% RAC increase in CT (statistically significant), and a 0.8% dice and 1.1% RAC increase in MRI (not statistically significant) in volume-wise acquisition. Our ablation study indicates that combining density and diversity criteria enhances the efficiency of BAL in musculoskeletal segmentation compared to using either criterion alone. Our sampling method is proven efficient in reducing annotation costs in image segmentation tasks. The combination of the proposed method and our BAL framework provides a semi-automatic way for efficient annotation of medical image datasets.

Refined parallel adaptive Bayesian quadrature for estimating small failure probabilities

Bayesian active learning methods have emerged for structural reliability analysis, showcasing more attractive features compared to existing active learning methods. The parallel adaptive Bayesian quadrature (PABQ) method, as a representative of them, allows to efficiently assessing small failure probabilities but faces the problem of empirically specifying several important parameters. The unreasonable parameter settings could lead to the inaccurate estimates of failure probability or the non-convergence of active learning. This study proposes a refined PABQ (R-PABQ) method by presenting three novel refinements to overcome the drawbacks of PABQ. Firstly, a sequential population enrichment strategy is presented and embedded into the importance ball sampling technique to solve the computer memory problem when involving large sample population. Secondly, an adaptive determination strategy for radius is developed to automatically adjust the sampling region during the active learning procedure. Lastly, an adaptive multi-point selection method is proposed to identify a batch of points to enable parallel computing. The effectiveness of the proposed R-PABQ method is demonstrated by four numerical examples. Results show that the proposed method can estimate small failure probabilities (e.g., 10−7∼10−9) with superior accuracy and efficiency over several existing active learning reliability methods.

A sequential sampling-based Bayesian numerical method for reliability-based design optimization

For efficiently solving the Reliability-Based Design Optimization (RBDO) problem with multi-modal, highly nonlinear and expensive-to-evaluate limit state functions (LSFs), a sequential sampling-based Bayesian active learning method is developed in this work. The penalty function method is embedded to transform the constrained optimization problem into a non-constrained one to reduce the model complexity. The proposed method for solving RBDO problems starts by training a Gaussian process (GP) model, in the augmented space of random and design variables. It is then based on an efficient sampling scheme for simulating the GP model, the adaptive Bayesian optimization (BO) and Bayesian reliability analysis (BRA) procedures are combined in a collaborative way for sequentially producing the joint training points. BO driven by expected improvement (EI) function is used for inferring the global optimum in the design space with global convergence, and the BRA equipped with U function is implemented for inferring the failure probabilities at the identified design points with the desired accuracy. The superiority of the proposed method is demonstrated with two numerical and two real-world engineering examples.

First principles molecular dynamics simulation and thermal decomposition kinetics study of CL-20.

2,4,6,8,10, 12-hexanitro-2,4,6,8,10, 12-hexazepane (CL-20) is a new energetic material with high performance and low sensitivity. In-depth study of the thermal decomposition mechanism of CL-20 is a necessary condition to improve its performance, ensure its safety, and optimize its application. On the basis of a large number of empirical force fields used in molecular dynamics simulation in the past, the machine learning augmented first-principles molecular dynamics method was used for the first time to simulate the thermal decomposition reaction of CL-20 at 2200 K, 2500 K, 2800 K, and 3000 K isothermal temperature. The main stable resulting compounds are N2, CO2, CO, H2O, andH2, where CO2 and H2O continue to decompose at higher temperatures. The initial decomposition pathways are denitration by N-N fracture, ring-opening by C-N bond fracture, and redox reaction involving NO2 and CL-20. After ring opening, two main compounds, fused tricyclic pyrazine and azadicyclic, were formed, which were decomposed continuously to form monocyclic pyrazine and pyrazole ring structures. The most common fragments formed during decomposition are those containing two, three, four, and six carbons. The formation rule and quantity of main small molecule intermediates and resulting stable products under different simulated temperatures were analyzed. Based on ab initio Bayesian active learning algorithm, efficient and accurate prediction of CL-20 is made using the dynamic machine learning function of Vienna Ab-Initio Simulation Package (VASP), which constructs the energy potential surface by learning a large number of data based on AIMD calculations. The result is a machine learning force field (MLFF). Then the molecular dynamics of CL-20 was simulated using the trained MLFF model. PAW pseudopotentials and generalized gradient approximation (GGA), namely, Perdew-Burke-Ernzerhof (PBE) functional, are used in the calculation. The plane wave truncation energy (ENCUT) is set to 550 eV, and using the Gaussian broadening, the thermal broadening size of the single-electron orbital is 0.05 eV. A van der Waals revision of the system with Grimme Version 3. The energy convergence accuracy (EDIFF) of electron self-consistent iteration is set to 1E-5 eV and 1E-6 eV, respectively. The two-step structure optimization is carried out using 1'1'1 k point grid and conjugate gradient method. The ENCUT was changed to 500 eV and EDIFF to 1E-5 eV, and NVT integration (ISIF = 2) of Langevin thermostat was used for machine learning force field training and AIMD simulation of the system.

Contrast response function estimation with nonparametric Bayesian active learning.

Multidimensional psychometric functions can typically be estimated nonparametrically for greater accuracy or parametrically for greater efficiency. By recasting the estimation problem from regression to classification, however, powerful machine learning tools can be leveraged to provide an adjustable balance between accuracy and efficiency. Contrast sensitivity functions (CSFs) are behaviorally estimated curves that provide insight into both peripheral and central visual function. Because estimation can be impractically long, current clinical workflows must make compromises such as limited sampling across spatial frequency or strong assumptions on CSF shape. This article describes the development of the machine learning contrast response function (MLCRF) estimator, which quantifies the expected probability of success in performing a contrast detection or discrimination task. A machine learning CSF can then be derived from the MLCRF. Using simulated eyes created from canonical CSF curves and actual human contrast response data, the accuracy and efficiency of the machine learning contrast sensitivity function (MLCSF) was evaluated to determine its potential utility for research and clinical applications. With stimuli selected randomly, the MLCSF estimator converged slowly toward ground truth. With optimal stimulus selection via Bayesian active learning, convergence was nearly an order of magnitude faster, requiring only tens of stimuli to achieve reasonable estimates. Inclusion of an informative prior provided no consistent advantage to the estimator as configured. MLCSF achieved efficiencies on par with quickCSF, a conventional parametric estimator, but with systematically higher accuracy. Because MLCSF design allows accuracy to be traded off against efficiency, it should be explored further to uncover its full potential.

Bayesian Active Learning for Sample Efficient 5G Radio Map Reconstruction

Bayesian Active Learning for Sample Efficient 5G Radio Map Reconstruction

Bayesian active learning isolation forest (B-ALIF): A weakly supervised strategy for anomaly detection

Due to the high cost of labelling large datasets, unsupervised Anomaly Detection models have been widely applied in many applications. In many scenarios, Anomaly Detection is typically embedded in Decision Support Systems, where users can monitor complex systems with increased efficiency. However, due to the subjective or domain-specific definition of anomalies, unsupervised models tend to perform poorly as they are based on general definitions of anomaly that might not coincide with the end-user expectations. In many cases human tagging of small amount of data can be feasible: For such scenarios, strategies to properly embed such information in the model and to guide the tagging may be extremely useful. Indeed to solve this problem, scientists are proposing models able to interactively query the end-user, training the detector towards the expected anomaly definition while reducing the human labelling effort; such scenario is especially feasible when Decision Support Systems are in place, since users have simple ways to interact with the Anomaly Detection modules. In this context, a novel Active Learning algorithm is here proposed, called Bayesian Active Learning Isolation Forest (B-ALIF), that, starting from the solution provided by a popular detection model named Isolation Forest, it is able to query the user and update by means of a Bayesian strategy. B-ALIF employs the unsupervised solution provided by the Isolation Forest as its prior, and then improves its detection capabilities as the user interacts with it. Experimental results on real datasets show that, with a limited human effort, the proposed approach is able to learn faster than other state-of-art weakly supervised and supervised approaches.