Bayesian Active Learning Research Articles

Bayesian active learning for multi‐objective feasible region identification in microwave devices

In microwave device and circuit design, many simulations are often needed to find a set of designs that satisfy one or multiple specifications chosen by the designer upfront: the feasible region. A novel Bayesian active learning framework is presented to accurately identify the feasible region with a low number of simulations. The technique leverages on a stochastic model to obtain an efficient and automated procedure. A suitable application example validates the proposed technique and shows its effectiveness to rapidly obtain many suitable designs.

Bounds optimization of model response moments: a twin-engine Bayesian active learning method

The efficient propagation of imprecise probabilities through expensive simulators has emerged to be one of the great challenges for mixed uncertainty quantification in computational mechanics. An active learning method, named Collaborative and Adaptive Bayesian Optimization (CABO), is developed for tackling this challenge by combining Bayesian Probabilistic Optimization and Bayesian Probabilistic Integration. Two learning functions are introduced as engines for CABO, where one is introduced for realizing the adaptive optimization search in the epistemic uncertainty space, and the other one is developed for adaptive integration in the aleatory uncertainty space. These two engines work in a collaborative way to create optimal design points adaptively in the joint uncertainty space, by which a Gaussian process regression model is trained and updated to approach the bounds of model response moments with pre-specified error tolerances. The effectiveness of CABO is demonstrated using a numerical example and two engineering benchmarks.

Accelerated Discovery of Zeolite Structures with Superior Mechanical Properties via Active Learning.

A Bayesian active learning platform is developed for the accelerated discovery of mechanically superior zeolite structures from more than half a million hypothetical candidates. An initial database containing the mechanical properties of synthesizable zeolites is trained to develop the machine learning regression model. Then, a Bayesian optimization scheme is implemented to identify zeolites with potentially excellent mechanical properties. The newly accumulated database consists of 871 labeled structures, and the uncertainty of the predictive model is reduced by 40% and 58% for the bulk and shear moduli, respectively. The model convergence shows that no further improvement occurs after the 10th iteration of optimizations. The proposed platform is able to discover 23 new zeolite structures that have unprecedented shear moduli; in one case, the shear modulus (127.81 GPa) is 250% higher than the previous data set. The proposed platform accelerates the material discovery process while maximizing computational efficiency and enhancing the predictive accuracy.

How useful is active learning for image‐based plant phenotyping?

AbstractDeep learning models have been successfully deployed for a diverse array of image‐based plant phenotyping applications including disease detection and classification. However, successful deployment of supervised deep learning models requires large amount of labeled data, which is a significant challenge in plant sciences (and most biological) domain due to the inherent complexities. Specifically, data annotation is costly, laborious, time consuming and needs domain expertise for phenotyping tasks, especially for diseases. To overcome this challenge, active learning algorithms have been proposed to reduce the amount of labeling needed by deep learning models to achieve good predictive performance. Active learning methods work by adaptively suggesting samples to annotate using an acquisition function to achieve maximum (classification) performance under a fixed labeling budget. We report the performance of four different active learning methods, (1) Deep Bayesian Active Learning (DBAL), (2) Entropy, (3) Least Confidence, and (4) core‐set, with conventional random sampling‐based annotation for two vastly different image‐based classification datasets. The first image dataset consists of soybean [Glycine max L. (Merr.)] leaves belonging to eight different soybean stresses and a healthy class, and the second consists of nine different weed species from the field. For a fixed labeling budget, we observed that the classification performance of deep learning models using active learning based acquisition strategies is better than random sampling‐based acquisition for both datasets. The integration of active learning strategies for data annotation can help mitigate labelling challenges in the plant sciences applications particularly where resources dedicated to annotations are limited.

High-efficient ab initio Bayesian active learning method and applications in prediction of two-dimensional functional materials.

Beyond the conventional trial-and-error method, machine learning offers a great opportunity to accelerate the discovery of functional materials, but still often suffers from difficulties such as limited materials data and the unbalanced distribution of target properties. Here, we propose the ab initio Bayesian active learning method that combines active learning and high-throughput ab initio calculations to accelerate the prediction of desired functional materials with ultrahigh efficiency and accuracy. We apply it as an instance to a large family (3119) of two-dimensional hexagonal binary compounds with unbalanced materials properties, and accurately screen out the materials with maximal electric polarization and proper photovoltaic band gaps, respectively, whereas the computational costs are significantly reduced by only calculating a few tenths of the possible candidates in comparison with a random search. This approach shows the enormous advantages for the cases with unbalanced distribution of target properties. It can be readily applied to seek a broad range of advanced materials.

On-the-fly closed-loop materials discovery via Bayesian active learning

Active learning—the field of machine learning (ML) dedicated to optimal experiment design—has played a part in science as far back as the 18th century when Laplace used it to guide his discovery of celestial mechanics. In this work, we focus a closed-loop, active learning-driven autonomous system on another major challenge, the discovery of advanced materials against the exceedingly complex synthesis-processes-structure-property landscape. We demonstrate an autonomous materials discovery methodology for functional inorganic compounds which allow scientists to fail smarter, learn faster, and spend less resources in their studies, while simultaneously improving trust in scientific results and machine learning tools. This robot science enables science-over-the-network, reducing the economic impact of scientists being physically separated from their labs. The real-time closed-loop, autonomous system for materials exploration and optimization (CAMEO) is implemented at the synchrotron beamline to accelerate the interconnected tasks of phase mapping and property optimization, with each cycle taking seconds to minutes. We also demonstrate an embodiment of human-machine interaction, where human-in-the-loop is called to play a contributing role within each cycle. This work has resulted in the discovery of a novel epitaxial nanocomposite phase-change memory material.

A detection method of worst conditions in vehicle dynamics tests using Bayesian Active Learning

A detection method of worst conditions in vehicle dynamics tests using Bayesian Active Learning

Bayesian3 Active Learning for the Gaussian Process Emulator Using Information Theory.

Gaussian process emulators (GPE) are a machine learning approach that replicates computational demanding models using training runs of that model. Constructing such a surrogate is very challenging and, in the context of Bayesian inference, the training runs should be well invested. The current paper offers a fully Bayesian view on GPEs for Bayesian inference accompanied by Bayesian active learning (BAL). We introduce three BAL strategies that adaptively identify training sets for the GPE using information-theoretic arguments. The first strategy relies on Bayesian model evidence that indicates the GPE’s quality of matching the measurement data, the second strategy is based on relative entropy that indicates the relative information gain for the GPE, and the third is founded on information entropy that indicates the missing information in the GPE. We illustrate the performance of our three strategies using analytical- and carbon-dioxide benchmarks. The paper shows evidence of convergence against a reference solution and demonstrates quantification of post-calibration uncertainty by comparing the introduced three strategies. We conclude that Bayesian model evidence-based and relative entropy-based strategies outperform the entropy-based strategy because the latter can be misleading during the BAL. The relative entropy-based strategy demonstrates superior performance to the Bayesian model evidence-based strategy.

Bayesian Active Learning for Optimization and Uncertainty Quantification in Protein Docking.

Ab initio protein docking represents a major challenge for optimizing a noisy and costly "black box"-like function in a high-dimensional space. Despite progress in this field, there is a lack of rigorous uncertainty quantification (UQ). To fill the gap, we introduce a novel algorithm, Bayesian active learning (BAL), for optimization and UQ of such black-box functions with applications to flexible protein docking. BAL directly models the posterior distribution of the global optimum (i.e., native structures) with active sampling and posterior estimation iteratively feeding each other. Furthermore, it uses complex normal modes to span a homogeneous, Euclidean conformation space suitable for high-dimensional optimization and constructs funnel-like energy models for quality estimation of encounter complexes. Over a protein-docking benchmark set and a CAPRI set including homology docking, we establish that BAL significantly improves against starting points from rigid docking and refinements by particle swarm optimization, providing a top-3 near-native prediction for one third targets. Quality assessment empowered with UQ leads to tight quality intervals with half range around 25% of the actual interface root-mean-square deviation and confidence level at 85%. BAL's estimated probability of a prediction being near-native achieves binary classification AUROC at 0.93 and area under the precision recall curve over 0.60 (compared to 0.50 and 0.14, respectively, by chance), which also improves ranking predictions. This study represents the first UQ solution for protein docking, with rigorous theoretical frameworks and comprehensive empirical assessments.

(Invited) Understanding Dynamical Stability of Batteries and Catalysts with First-Principles and Ex-machina Computations

Electrochemical stability windows of electrolytes largely determine the limitations of operating regimes of lithium-ion batteries, but the degradation mechanisms are difficult to characterize and poorly understood. Using computational quantum chemistry to investigate the oxidative decomposition that govern voltage stability of multi-component organic electrolytes, we find that electrolyte decomposition is a process involving the solvent and the salt anion and requires explicit treatment of their coupling. We find that the ionization potential of the solvent-anion system is often lower than that of the isolated solvent or the anion. This understanding of the oxidation mechanism allows to formulate a simple predictive model that explains experimentally observed trends in the onset voltages of degradation of electrolytes near the cathode. This model opens opportunities for rapid rational design of stable electrolytes for high-energy batteries.Design of materials based on dynamical and transport properties, important for batteries and catalysts, involves systematic exploration of many structural and compositional variations and hence requires very fast and accurate computations of dynamical phenomena. I will present recent progress and challenges in “ex-machina” materials design, a new paradigm in which an automated closed-loop machine learning algorithm constructs non-parametric Bayesian force fields that combine first-principles accuracy with internal quantitative uncertainty and prior information of physical symmetries. We apply this method, implemented in the FLARE framework [1], to large-scale dynamics simulations of alloys, ion conductors, catalysts and 2D materials.[1] J. Vandermause, S. B. Torrisi, S. Batzner, A. M. Kolpak, B. Kozinsky, “On-the-Fly Bayesian Active Learning of Interpretable Force-Fields for Atomistic Rare Events”, (2019), arXiv:1904.02042

Discovering Causal Structures in Bayesian Gaussian Directed Acyclic Graph Models

SummaryCausal directed acyclic graphs (DAGs) are naturally tailored to represent biological signalling pathways. However, a causal DAG is only identifiable up to Markov equivalence if only observational data are available. Interventional data, based on exogenous perturbations of the system, can greatly improve identifiability. Since the gain of an intervention crucially depends on the intervened variables, a natural issue is devising efficient strategies for optimal causal discovery. We present a Bayesian active learning procedure for Gaussian DAGs which requires no subjective specification on the side of the user, explicitly takes into account the uncertainty on the space of equivalence classes (through the posterior distribution) and sequentially proposes the choice of the optimal intervention variable. In simulation experiments our method, besides surpassing designs based on a random choice of intervention nodes, shows decisive improvements over currently available algorithms and is competitive with the best alternative benchmarks. An important reason behind this strong performance is that, unlike non-Bayesian algorithms, our utility function naturally incorporates graph estimation uncertainty through the posterior edge inclusion probability. We also reanalyse the Sachs data on protein signalling pathways from an active learning perspective and show that DAG identification can be achieved by using only a subset of the available intervention samples.

Bayesian Active Learning for Choice Models With Deep Gaussian Processes

This paper proposes an active learning algorithm and models which can gradually learn individual's preference through pairwise comparisons. The active learning scheme aims at finding individual's most preferred choice (e.g. an airline itinerary) with minimized number of pairwise comparisons. The pairwise comparisons are encoded into probabilistic models based on assumptions of choice models and deep Gaussian processes. More specifically, this paper develops two novel probabilistic models assuming correlated Gumbel noises and latent utility functions. One is based on shallow Gaussian priors and the other assumes deep Gaussian priors. In the active learning algorithm, the next-to-compare decision is determined by an original acquisition function. The proposed algorithm and models have been benchmarked using functions with multiple local optima and one public airline itinerary dataset. In both experiments, nests are designed to capture correlated Gumbel noises. The experiments indicate the effectiveness of our active learning algorithm and models. The deep Gaussian models are proven to find the best choice with a lower number of pairwise comparisons than the shallow one. In both experiments, deep Gaussian models outperform the shallow model. The shallow model is recommended when the choice set is large and less computational time is required (e.g. in one experiment, deep Gaussian models require approximately 60%-70% more time on average).

Efficient Design of Engine Mount Characteristics Using Feature Extraction to Predict Vehicle Vibration and Bayesian Active Learning

Efficient Design of Engine Mount Characteristics Using Feature Extraction to Predict Vehicle Vibration and Bayesian Active Learning

Fast estimation of equal-loudness contours using Bayesian active learning and direct scaling

Fast estimation of equal-loudness contours using Bayesian active learning and direct scaling

Application of Bayesian Active Learning to the Estimation of Auditory Filter Shapes Using the Notched-Noise Method.

Time-efficient hearing tests are important in both clinical practice and research studies. This particularly applies to notched-noise tests, which are rarely done in clinical practice because of the time required. Auditory-filter shapes derived from notched-noise data may be useful for diagnosis of the cause of hearing loss and for fitting of hearing aids, especially if measured over a wide range of center frequencies. To reduce the testing time, we applied Bayesian active learning (BAL) to the notched-noise test, picking the most informative stimulus parameters for each trial based on nine Gaussian Processes. A total of 11 hearing-impaired subjects were tested. In 20 to 30 min, the test provided estimates of signal threshold as a continuous function of frequency from 500 to 4000 Hz for nine notch widths and for notches placed both symmetrically and asymmetrically around the signal frequency. The thresholds were found to be consistent with those obtained using a 2-up/1-down forced-choice procedure at a single center frequency. In particular, differences in threshold between the methods did not vary with notch width. An independent second run of the BAL test for one notch width showed that it is reliable. The data derived from the BAL test were used to estimate auditory-filter width and asymmetry and detection efficiency for center frequencies from 500 to 4000 Hz. The results agreed with expectations for cochlear hearing losses that were derived from the audiogram and a hearing model.

Bayesian Active Learning for Collaborative Task Specification Using Equivalence Regions

Specifying complex task behaviours while ensuring good robot performance may be difficult for untrained users. We study a framework for users to specify rules for acceptable behaviour in a shared environment such as industrial facilities. As non-expert users might have little intuition about how their specification impacts the robot's performance, we design a learning system that interacts with the user to find an optimal solution. Using active preference learning, we iteratively show alternative paths that the robot could take on an interface. From the user feedback ranking the alternatives, we learn about the weights that users place on each part of their specification. We extend the user model from our previous work to a discrete Bayesian learning model and introduce a greedy algorithm for proposing alternative that operates on the notion of equivalence regions of user weights. We prove that with this algorithm the revision active learning process converges on the user-optimal path. In simulations on realistic industrial environments, we demonstrate the convergence and robustness of our approach.

Audiogram estimation using Bayesian active learning.

Two methods for estimating audiograms quickly and accurately using Bayesian active learning were developed and evaluated. Both methods provided an estimate of threshold as a continuous function of frequency. For one method, six successive tones with decreasing levels were presented on each trial and the task was to count the number of tones heard. A Gaussian Process was used for classification and maximum-information sampling to determine the frequency and levels of the stimuli for the next trial. The other method was based on a published method using a Yes/No task but extended to account for lapses. The obtained audiograms were compared to conventional audiograms for 40 ears, 19 of which were hearing impaired. The threshold estimates for the active-learning methods were systematically from 2 to 4 dB below (better than) those for the conventional audiograms, which may indicate a less conservative response criterion (a greater willingness to respond for a given amount of sensory information). Both active-learning methods were able to allow for wrong button presses (due to lapses of attention) and provided accurate audiogram estimates in less than 50 trials or 4 min. For a given level of accuracy, the counting task was slightly quicker than the Yes/No task.

A Bayesian Active Learning Experimental Design for Inferring Signaling Networks.

Machine learning methods for learning network structure are applied to quantitative proteomics experiments and reverse-engineer intracellular signal transduction networks. They provide insight into the rewiring of signaling within the context of a disease or a phenotype. To learn the causal patterns of influence between proteins in the network, the methods require experiments that include targeted interventions that fix the activity of specific proteins. However, the interventions are costly and add experimental complexity. We describe an active learning strategy for selecting optimal interventions. Our approach takes as inputs pathway databases and historic data sets, expresses them in form of prior probability distributions on network structures, and selects interventions that maximize their expected contribution to structure learning. Evaluations on simulated and real data show that the strategy reduces the detection error of validated edges as compared with an unguided choice of interventions and avoids redundant interventions, thereby increasing the effectiveness of the experiment.

Constrained Bayesian Active Learning of Interference Channels in Cognitive Radio Networks

In this paper, a sequential probing method for interference constraint learning is proposed to allow a centralized cognitive radio network (CRN) accessing the frequency band of a primary user (PU) in an underlay cognitive scenario with a designed PU protection specification. The main idea is that the CRN probes the PU, and subsequently, eavesdrops the reverse PU link to acquire the binary ACK/NACK packet. This feedback indicates whether the probing-induced interference is harmful or not and can be used to learn the PU interference constraint. The cognitive part of this sequential probing process is the selection of the power levels of the secondary users that aims to learn the PU interference constraint with a minimum number of probing attempts while setting a limit on the number of harmful probing-induced interference events or equivalently of NACK packet observations over a time window. This constrained design problem is studied within the active learning (AL) framework and an optimal solution is derived and implemented with a sophisticated, accurate, and fast Bayesian learning method, the expectation propagation. The performance of this solution is also demonstrated through numerical simulations and compared with modified versions of AL techniques we developed in earlier work.

A fast test for determining the edge frequency of a dead region

The presence and edge frequency, fe, of a dead region in the cochlea can be diagnosed using psychophysical tuning curves (PTCs). When the signal frequency, fs, falls in a dead region, the tip of the PTC lies close to fe, rather than close to fs. However, measurement of PTCs is time consuming, limiting their application in clinical practice. We have developed a fast test based on Bayesian active learning. Instead of estimating an entire PTC, we estimate parameters of an individual hearing model, including fe. The task is to detect a fixed signal in the presence of a masker whose level and frequency vary across trials. After each trial, the next masker level and frequency are chosen to produce maximum reduction of the uncertainty about the parameters. The results for four participants tested so far were close to those obtained using “fast” PTCs. The Bayesian procedure has two advantages compared to PTCs: it allows quantification of the reliability of the subjects, estimated from the standard deviation of a cumulative Gaussian fitted to the psychometric function; and masker levels and frequencies can be restricted to being close to the estimated minimum of the PTC, avoiding unnecessary presentation of high level sounds.