Active Learning Framework Research Articles

Active learning-driven uncertainty reduction for in-flight particle characteristics of atmospheric plasma spraying of silicon

The first-of-its-kind use of the active learning (AL) framework in thermal spray is adapted to enhance the prediction accuracy of the in-flight particle characteristics. The successful AL framework implementation via Bayesian Optimisation is beneficial in, first, reducing the maximum uncertainty, which greatly improves the prediction accuracy and informativeness of the existing database. Second, it reduces local uncertainty around a contrived test point that offers the capability to find improvement in a limited search area, allowing an accurate prediction around a desired test point. The dataset for Machine Learning (ML) training consists of 26 atmospheric plasma spray (APS) parameters of silicon and a further six AL-guided spray runs carried out to reduce maximum uncertainty in the initial database. On average, a 52.9% improvement (error reduction) of RMSE and an R2 increase of 8.5% were reported on the predicted in-flight particle velocities and temperatures after the AL-driven optimisation. Furthermore, the contrived test point optimisation to predict the best possible characteristics in a limited search space resulted in a three-fold increase in prediction accuracy compared to the non-optimised prediction. The AL-driven optimisation proved to be greatly beneficial for resource-intensive thermal spraying, as the framework not only allowed an accurate prediction of the in-flight particle characteristics but also found expected improvement around a desired in-flight characteristic. Furthermore, the framework uses the Gaussian Process (GP) ML model as a surrogate that generalises a global solution without necessarily involving physical and underlying mechanisms, thus extending the framework to other thermal spraying methods.

Prediction of equivalent sand-grain size and identification of drag-relevant scales of roughness – a data-driven approach

Despite decades of research, a universal method for prediction of roughness-induced skin friction in a turbulent flow over an arbitrary rough surface is still elusive. The purpose of the present work is to examine two possibilities; first, predicting equivalent sand-grain roughness size $k_s$ based on the roughness height probability density function and power spectrum (PS) leveraging machine learning as a regression tool; and second, extracting information about relevance of different roughness scales to skin-friction drag by interpreting the output of the trained data-driven model. The model is an ensemble neural network (ENN) consisting of 50 deep neural networks. The data for the training of the model are obtained from direct numerical simulations (DNS) of turbulent flow in plane channels over 85 irregular multi-scale roughness samples at friction Reynolds number $Re_\tau =800$ . The 85 roughness samples are selected from a repository of 4200 samples, covering a wide parameter space, through an active learning (AL) framework. The selection is made in several iterations, based on the informativeness of samples in the repository, quantified by the variance of ENN predictions. This AL framework aims to maximize the generalizability of the predictions with a certain amount of data. This is examined using three different testing data sets with different types of roughness, including 21 surfaces from the literature. The model yields overall mean error 5 %–10 % on different testing data sets. Subsequently, a data interpretation technique, known as layer-wise relevance propagation, is applied to measure the contributions of different roughness wavelengths to the predicted $k_s$ . High-pass filtering is then applied to the roughness PS to exclude the wavenumbers identified as drag-irrelevant. The filtered rough surfaces are investigated using DNS, and it is demonstrated that despite significant impact of filtering on the roughness topographical appearance and statistics, the skin-friction coefficient of the original roughness is preserved successfully.

Leveraging on Virtual Laboratory-Based Instruction to Achieve Active Classroom Interaction in Teaching and Learning of Physics in Secondary Schools in Kenya

The philosophy that advances the study of physics as one of the learning areas in secondary schools cannot be overlooked due to the realization of the contributions of physics towards the industrial and vocational development of the country. The major concern of this particular study is the consistent poor average scores in the subject currently witnessed at the Kenya Certificate of Secondary Education (KCSE) level in Physics, which can be attributed to low student motivation and traditional teaching strategies. For instance, in the years 2016, 2017, 2018, and 2019, Kisumu County registered low mean scores of 4.23, 4.98, 4.67, and 4.10, respectively, in physics in KCSE. The use of Information and Communications Technology (ICT) in the teaching and learning of physics can be used to promote learners’ centered teaching and learning of physics and eventually improve physics scores. This study therefore underscores the need for virtual laboratory-based instruction (VLBI) to realize active classroom interaction during the teaching and learning of physics. Technology-based enhanced learning leverages all learners, irrespective of their economic and demographic background. The study therefore aimed at establishing the effect of VLBI on students’ level of interaction in the classroom in learning physics in secondary schools in Kenya. The study is supported by behaviorism learning theory and adopted quasi-experimental research designs. The physics teachers were purposefully sampled from each selected school. Physics teachers were purposefully sampled from each of the schools selected. There were 358 students and 72 teachers in the sample. The data was analyzed using both descriptive and inferential statistics. The study concluded that virtual laboratories enhanced the use of experimental teaching approaches and that there was no statistical significance between the knowledge of the teacher about the selected ICT framework and the use of virtual experiments. There is a need to develop a prototype for a technology-enriched ICT framework for active learning of physics that will provide teachers and students with an interactive platform to promote active physics learning.

MCDAL: Maximum Classifier Discrepancy for Active Learning.

Recent state-of-the-art active learning methods have mostly leveraged generative adversarial networks (GANs) for sample acquisition; however, GAN is usually known to suffer from instability and sensitivity to hyperparameters. In contrast to these methods, in this article, we propose a novel active learning framework that we call Maximum Classifier Discrepancy for Active Learning (MCDAL) that takes the prediction discrepancies between multiple classifiers. In particular, we utilize two auxiliary classification layers that learn tighter decision boundaries by maximizing the discrepancies among them. Intuitively, the discrepancies in the auxiliary classification layers' predictions indicate the uncertainty in the prediction. In this regard, we propose a novel method to leverage the classifier discrepancies for the acquisition function for active learning. We also provide an interpretation of our idea in relation to existing GAN-based active learning methods and domain adaptation frameworks. Moreover, we empirically demonstrate the utility of our approach where the performance of our approach exceeds the state-of-the-art methods on several image classification and semantic segmentation datasets in active learning setups.

Bidirectional Active Transfer Learning for Adaptive Power System Stability Assessment and Dominant Instability Mode Identification

Deep learning (DL) is a useful tool for power system stability assessment (PSSA) and dominant instability mode (DIM) identification. However, when faced with operational variability, the performance of DL models degrades. This paper proposes a bidirectional active transfer learning (Bi-ATL) framework for more adaptive PSSA and DIM identification, where the DL model is easier to adapt to unlearned operating conditions with fewer newly labeled instances. At the instance level, forward active learning and backward active learning are integrated to progressively build a mixed instance set by actively including the most label-worthy instances of new operating conditions and actively eliminating the most useless original operating condition instances. Then at the model parameter level, the mixed instance set is utilized to fine-tune the original DL model to new operating conditions. The Bi-ATL framework synthesizes three-way information of the instances and model of the original operating condition, and a few labeled instances of new operating conditions for more efficient adaptation. Intensive case studies conducted on a benchmark power system (CEPRI 36-bus system) and a real-world large-scale power system (Northeast China Power System-2131 bus) validate the efficacy and efficiency of the Bi-ATL framework as well as the role of the three-way information.

Reduced order modeling of hydrodynamic interactions between a submarine and unmanned underwater vehicle using non-myopic multi-fidelity active learning

Several efforts have been dedicated to developing computational tools capable of predicting the hydrodynamic forces and moments of Unmanned Underwater Vehicles (UUVs). However, there is no method at the moment that allows for real-time computational modeling of all the complex hydrodynamic interaction forces and moments that a UUV experiences when operating in close proximity to a moving submarine. Real-time modeling of these hydrodynamic interactions is essential to simulate the motion required to launch and recover UUVs from submarines. Potential flow models are often fast enough to be used in real time, but lack the accuracy of Computational Fluid Dynamics (CFD) simulations, which often take hours or days to solve. Here, we formulate the problem in the context of machine learning, specifically active learning. The goal is to develop a surrogate model capable of predicting the UUV and submarine hydrodynamic interactions in real time using a very small number of carefully selected CFD simulations. We introduce a new active learning framework called Non-Myopic Multi-Fidelity Active Learning for Gaussian Process (GP) regression that accelerates the convergence of the surrogate model by utilizing the low cost of the low fidelity simulations to explore the domain, as well as optimally selected high fidelity simulations to improve the model accuracy. The resulting surrogate model can be integrated into UUV control and autonomy systems and motion simulators to further enable UUV launch and recovery from submarines. This new active learning method may also be used to create higher accuracy and lower cost surrogate models in other real world applications.

Classifying and modeling secondary students’ active learning in a virtual learning environment through generated questions

The present research study investigated students' active learning activities at three time points during the use of a virtual learning environment (VLE) on cardiac anatomy and physiology. Because VLEs are not only immersive but also interactive, the user must engage with the instruction to move the instruction forward; thus, engaging the user in an active (rather than passive) educative experience. What types of active learning is garnered from such educational technologies, among secondary aged science learners, was the focus of the study. By using open-ended questioning and the active learning framework by Chi (2009), this research study collected students' questions they had about the science content which was then categorized inductively and deductively by active processes (activation of prior knowledge), constructive processes (garnering new knowledge), and interactive processes (creating new understandings beyond the scope of the curriculum). Results from coding questions from 151 secondary students (75 in sixth grade and 76 in ninth grade) revealed significance in sixth graders’ abilities to progress from active processes to the higher and more sophisticated learning processes; significant changes were found among ninth graders to progress from active to constructive processes and from constructive to interactive learning processes. The presence of interactive processes suggests that VLEs are immersive, interactive, and able to foster robust science learning.

Cluster-MLP: An Active Learning Genetic Algorithm Framework for Accelerated Discovery of Global Minimum Configurations of Pure and Alloyed Nanoclusters.

Structural characterization of nanoclusters is one of the major challenges in nanocluster modeling owing to the multitude of possible configurations of arrangement of cluster atoms. The genetic algorithm (GA), a class of evolutionary algorithms based on the principles of natural evolution, is a commonly employed search method for locating the global minimum configuration of nanoclusters. Although a GA search at the DFT level is required for the accurate description of a potential energy surface to arrive at the correct global minimum configuration of nanoclusters, computationally expensive DFT evaluation of the significantly larger number of cluster geometries limits its practicability. Recently, machine learning potentials (MLP) that are learned from DFT calculations gained significant attention as computationally cheap alternative options that provide DFT level accuracy. As the accuracy of the MLP predictions is dependent on the quality and quantity of the training DFT data, active learning (AL) strategies have gained significant momentum to bypass the need of large and representative training data. In this application note, we present Cluster-MLP, an on-the-fly active learning genetic algorithm framework that employs the Flare++ machine learning potential (MLP) for accelerating the GA search for global minima of pure and alloyed nanoclusters. We have used a modified version the Birmingham parallel genetic algorithm (BPGA) for the nanocluster GA search which is then incorporated into distributed evolutionary algorithms in Python (DEAP), an evolutionary computational framework for fast prototyping or technical experiments. We have shown that the incorporation of the AL framework in the BPGA significantly reduced the computationally expensive DFT calculations. Moreover, we have shown that both the AL-GA and DFT-GA predict the same global minima for all the clusters we tested.

An Online Active Broad Learning Approach for Real-Time Safety Assessment of Dynamic Systems in Nonstationary Environments.

Real-time safety assessment of the complex dynamic systems in nonstationary environments is of great significance for avoiding the potential hazards. In this case, the update procedure with high assessment accuracy and training speed is crucial and meaningful in the dynamic streaming setting. Generally, the performance of most online learning approaches will be negatively affected by limited annotated samples in such a setting. Moreover, the time cost of advanced conventional methods with retaining procedures is relatively high, constraining their practicality. In this article, a novel online active broad learning approach, termed OABL, is proposed. In detail, the effectiveness of the broad learning system in the framework of online active learning is first revealed and verified. A reasonable dynamic asymmetric query strategy is then designed with a limited annotation budget to actively annotate the relatively valuable samples, which is beneficial to mitigating the negative effects of class imbalance. In this context, the advantage of the human-in-the-loop characteristic is also effectively used to control the evolution direction of the learner during the incremental update, which makes it better able to adapt to complex and nonstationary environments. Several related experiments are conducted with the realistic data of JiaoLong deep-sea manned submersible. Results show the effectiveness and practicality of the proposal compared with the existing advanced approaches.

A feature-based physics-constrained active dictionary learning scheme for image-based additive manufacturing process monitoring

Compressed sensing is a new technique for data acquisition, which reconstructs the original signal from a small number of measurements. The performance of compressed sensing depends on the design of measurement and basis matrices. To improve the reconstruction accuracy, dictionary learning methods have been developed to customize basis matrices for specific signals. However, it may be difficult to obtain the theoretically optimal results because of the restrictions in physical realization such as the sensor placement and accessibility in the working environment. In this paper, a feature-based physics-constrained active dictionary learning (FB-PCADL) framework is proposed to solve the compression and classification problems simultaneously for image-based additive manufacturing process monitoring. A new active dictionary learning strategy is developed to optimize the measurement matrix, where the reconstruction errors are gradually reduced with the new measurements taken adaptively at the positions where the largest reconstruction errors occur. In addition, important features in images are identified and used to enhance the classification accuracy. The FB-PCADL framework is demonstrated with thermal images in monitoring fused filament fabrication process and optical images in monitoring laser powder bed fusion process.

Generative abstraction of Markov population processes

Markov population models are a widespread formalism used to model the dynamics of complex systems, with applications in systems biology and many other fields. The associated Markov stochastic process in continuous time is often analyzed by simulation, which can be costly for large or stiff systems, particularly when a massive number of simulations has to be performed, e.g. in a multi-scale model. A strategy to reduce computational load is to abstract the population model, replacing it with a simpler stochastic model, faster to simulate. Here we pursue this idea, exploring and comparing state-of-the-art generative models, which are flexible enough to automatically learn distributions over entire trajectories, rather than single simulation steps, from observed realizations of the system. In particular, we compare a Generative Adversarial setting with a Score-based Diffusion approach and show how the latter outperforms the former both in terms of accuracy and stability at the cost of slightly higher simulation times. To improve the accuracy of abstract samples, we develop an active learning framework to enrich our dataset with observations whose expected satisfaction of a temporal requirement differs significantly from the abstract one. We experimentally show how the proposed abstractions are well suited to work on multi-scale and data-driven scenarios, meaning that we can infer a (black-box) dynamical model from a pool of real data.

A Multi-objective Optimization Planning Framework for Active Distribution System Via Reinforcement Learning

The effective planning of active distribution networks is crucial for utility companies to make informed decisions regarding investments in distributed generation, reliability assessment, reactive power planning, substation revisions, and feeder repositioning. However, the dynamic nature of the solution space makes it challenging for model-based optimization methods to ensure computational performance in active distribution network planning. To address this issue, this study proposes a planning method that focuses on improving computational performance through the continuous updating of the planning model’s solution space during the reinforcement learning training process. Based on simulations conducted on the IEEE 33-bus test system, the proposed planning strategy successfully enhances computational performance while minimizing investment costs compared to other strategies. With the proposed method, the investment cost and the operation cost are reduced by 32.42% and 23.91%, respectively.

Efficient sampling-based Bayesian Active Learning for synaptic characterization.

Bayesian Active Learning (BAL) is an efficient framework for learning the parameters of a model, in which input stimuli are selected to maximize the mutual information between the observations and the unknown parameters. However, the applicability of BAL to experiments is limited as it requires performing high-dimensional integrations and optimizations in real time. Current methods are either too time consuming, or only applicable to specific models. Here, we propose an Efficient Sampling-Based Bayesian Active Learning (ESB-BAL) framework, which is efficient enough to be used in real-time biological experiments. We apply our method to the problem of estimating the parameters of a chemical synapse from the postsynaptic responses to evoked presynaptic action potentials. Using synthetic data and synaptic whole-cell patch-clamp recordings, we show that our method can improve the precision of model-based inferences, thereby paving the way towards more systematic and efficient experimental designs in physiology.

HAL-IA: A Hybrid Active Learning framework using Interactive Annotation for medical image segmentation.

High performance of deep learning models on medical image segmentation greatly relies on large amount of pixel-wise annotated data, yet annotations are costly to collect. How to obtain high accuracy segmentation labels of medical images with limited cost (e.g. time) becomes an urgent problem. Active learning can reduce the annotation cost of image segmentation, but it faces three challenges: the cold start problem, an effective sample selection strategy for segmentation task and the burden of manual annotation. In this work, we propose a Hybrid Active Learning framework using Interactive Annotation (HAL-IA) for medical image segmentation, which reduces the annotation cost both in decreasing the amount of the annotated images and simplifying the annotation process. Specifically, we propose a novel hybrid sample selection strategy to select the most valuable samples for segmentation model performance improvement. This strategy combines pixel entropy, regional consistency and image diversity to ensure that the selected samples have high uncertainty and diversity. In addition, we propose a warm-start initialization strategy to build the initial annotated dataset to avoid the cold-start problem. To simplify the manual annotation process, we propose an interactive annotation module with suggested superpixels to obtain pixel-wise label with several clicks. We validate our proposed framework with extensive segmentation experiments on four medical image datasets. Experimental results showed that the proposed framework achieves high accuracy pixel-wise annotations and models with less labeled data and fewer interactions, outperforming other state-of-the-art methods. Our method can help physicians efficiently obtain accurate medical image segmentation results for clinical analysis and diagnosis.

Development of a machine-learning interatomic potential for uranium under the moment tensor potential framework

Uranium is attracting growing interest across fundamental science and nuclear research, where atomistic simulations remain challenging. In this study, we developed a novel uranium interatomic potential based on a machine-learning approach embedded in the moment tensor potential package. An active learning framework was utilized in the potential training, using density-functional theory plus Hubbard U (DFT+U) modified data. The potential accurately reproduced the lattice parameters, cohesive energy, elastic, vibrational, and thermodynamic properties of α-U compared to DFT+U calculations and experimental results. The basic properties of other phases, including β, γ, HCP, and FCC phases, were also validated. In addition, molecular dynamics simulations were used to reproduce the temperature-induced allotropic transformation from α-U to γ-U, which agrees with experimental observations. Our potential provides a computationally efficient means to study the physical behavior of uranium with nearly DFT accuracy.

An active learning framework and assessment of inter-annotator agreement facilitate automated recogniser development for vocalisations of a rare species, the southern black-throated finch (Poephila cincta cincta)

The application of machine learning methods has led to major advances in the development of automated recognisers used to analyse bioacoustics data. To further improve the performance of automated call recognisers, we investigated the development of efficient data annotation strategies and how best to address uncertainty around ambiguous vocalisations. These challenges present a particular problem for species whose vocalisations are rare in field recordings, where collecting enough training data can be problematic and a species' vocalisations may be poorly documented.We provide an open access solution to address these challenges using two strategies. First, we applied an active learning framework to iteratively improve a convolutional neural network (CNN) model able to automate call identification for a target rare bird species, the southern black-throated finch (Poephila cincta cincta). We collected 9098 h of unlabelled audio recordings from a field study in the Desert Uplands Bioregion of Queensland, Australia, and used active learning to prioritise human annotation effort towards data that would best improve model fit. Second, we progressed methods for managing ambiguous vocalisations by applying machine learning methods more commonly used in medical image analysis and natural language processing. Specifically, we assessed agreement among human annotators and the CNN model (i.e. inter-annotator agreement) and used this to determine realistic performance outcomes for the CNN model and to identify areas where inter-annotator agreement may be improved. We also applied a classification approach that allowed the CNN model to classify sounds into an ‘uncertain’ category, which replicated a requirement of human-annotation and facilitated the comparison of human-model annotation performance.We found that active learning was an efficient strategy to build a CNN model where there was limited labelled training data available, and target calls were extremely rare in the unlabelled data. As few as five active learning iterations, generating a final labelled dataset of 1073 target calls and 5786 non-target sounds, were required to train a model to identify the target species with comparable performance to experts in the field.Assessment of inter-annotator agreement identified a bias in our model to align predictions most closely with those of the primary annotator and identified significant differences in inter-annotator agreement among subsets of our acoustic data. Our results highlight the use of inter-annotator agreement to understand model performance and identify areas for improvement in data annotation. We also show that excluding ambiguous vocalisations during data annotation results in an overestimation of model performance, an important consideration for datasets with inter-annotator disagreement.

Density-based one-shot active learning for image segmentation

Image segmentation is a key step in image processing tasks, which has significant applications in computer vision field such as medical image analysis, scene understanding and video monitoring, etc. However, image segmentation tasks usually require a large number of labeled samples to obtain great performance of convolutional neural networks (CNNs). Active learning (AL) can select valuable samples for annotation, so as to reduce the annotation cost as much as possible while maintaining the performance of CNNs. Further, one-shot AL can select valuable samples by once, which eliminates the need for iterative sample selection and annotation. However, existing one-shot AL approaches extremely rely on complex clustering algorithm, which brings a limitation in practice, i.e., we often do not know how to set the hyperparameters. In this paper, we propose a clustering-free one-shot AL framework, which is based on self-supervised feature learning and density-based query strategy. Our framework can select samples with high local density robustly against hyperparameters. The experimental results are impressive that state-of-the-art one-shot active learning performance can be achieved with simple density-based sampling.

Non data hungry smart composite manufacturing using active transfer learning with sigma point sampling (SPSATL)

As an emerging smart manufacturing paradigm, Industry 4.0 employs advancements in sensory technologies and cyber physical systems (CPS), to automate and optimize manufacturing processes. However, using conventional machine learning (ML) in this paradigm can be challenging especially in dynamic systems where materials used in the production plan can change constantly in order to adapt to the market needs. Another factor that currently hinders ML-based automation in many manufacturing processes is the lack of labeled data with usable features. In this paper, we address the aforementioned challenges in an aerospace composites manufacturing case study, by using an active transfer learning (ATL) framework. Active learning, in a semi-supervised setting, functions as a family of ML where the model is trained only on the most relevant labeled data. Transfer learning, on the other hand, is used to create robust ML models that reduce the effect of data distribution shifts in different source and target datasets. In particular, we present an enhanced approach, ATL using Sigma point sampling (SPSATL). SPSATL leverages representative subsets constructed with sigma points to capture true data distributions in a more concise and informative way, making AL procedures faster and more data-efficient. The proposed SPSATL approach is evaluated in a composites autoclave processing case study and the model outperformed the state-of-the-art sampling models in terms of accuracy (as high as 12% increase) and the volume of training data required (as low as 50% decrease).

Chatbots for active learning: A case of phishing email identification

Chatbots represent a promising approach to provide instructional content and facilitate active learning processes. However, there is a lack of knowledge as how to design chatbot interactions for active learning. In response to this knowledge gap, we conducted an experimental study (n = 164) comparing four modes for providing instructional content in chatbots, with varying demands for cognitive engagement. The four modes – passive, active, constructive, and interactive – were based on the ICAP framework of active learning. The learning content concerned identification of phishing emails and the four modes were distinguished by how the participants were invited to engage with the content during their chatbot interaction. The ICAP modes of higher cognitive engagement required participants to spend more time on the interaction and led to perceptions of higher subjective learning outcome. However, the effects of the different ICAP modes were not found to be significantly different in terms of user engagement, social presence, intention to use, or objective learning outcomes. The study represents an important first step towards understanding the design of chatbots for active learning.

Hierarchical active learning for defect localization in 3D systems

Aim: Advanced sensing and imaging is capable to retrieve rich information of complex systems, which can be integrated with underlying physics to develop a personalized simulation framework to discern system internal property for defect localization. Simulation-based defect localization involves modeling the heterogeneous 3D physical systems and calibrating spatial-varying model parameters. In this paper, we aim to develop an effective simulation-based active learning framework for defect localization in complex 3D systems. Methods: We develop a Hierarchical Gaussian Process (GP)-based Active Learning (HGPAL) framework to estimate the spatial-varying model parameters for reliable defect localization in complex 3D systems. We first develop a GP regression model on the 2D embedding of a 3D geometry to capture the geometric information. Second, we propose a Hierarchical GP-based active learning approach composed of master GP and sub-GP modeling to account for the system heterogeneity and further estimate the spatial-varying model parameter for defect localization. Results: We evaluate the performance of the proposed HGPAL framework in a 3D body-heart system to identify and localize infarct in the heart. Numerical experiments demonstrate the effectiveness of our HGPAL framework for 3D defect localization. Conclusion: This study successfully developed and tested an HGPAL framework for defect localization in complex 3D systems. The results suggest that this approach can efficiently and accurately localize defects in 3D systems, demonstrating its potential for practical applications in precision engineering system design and healthcare treatment planning.