Volume Of Training Data Research Articles

Height estimation from single aerial imagery using contrastive learning based multi-scale refinement network

ABSTRACT Height map estimation from a single aerial image plays a crucial role in localization, mapping, and 3D object detection. Deep convolutional neural networks have been used to predict height information from single-view remote sensing images, but these methods rely on large volumes of training data and often overlook geometric features present in orthographic images. To address these issues, this study proposes a gradient-based self-supervised learning network with momentum contrastive loss to extract geometric information from non-labeled images in the pretraining stage. Additionally, novel local implicit constraint layers are used at multiple decoding stages in the proposed supervised network to refine high-resolution features in height estimation. The structural-aware loss is also applied to improve the robustness of the network to positional shift and minor structural changes along the boundary area. Experimental evaluation on the ISPRS benchmark datasets shows that the proposed method outperforms other baseline networks, with minimum MAE and RMSE of 0.116 and 0.289 for the Vaihingen dataset and 0.077 and 0.481 for the Potsdam dataset, respectively. The proposed method also shows around threefold data efficiency improvements on the Potsdam dataset and domain generalization on the Enschede datasets. These results demonstrate the effectiveness of the proposed method in height map estimation from single-view remote sensing images.

Breast mass segmentation using mammographic data: a systematic review

ABSTRACT Breast cancer continues to be the most prevalent cancer infecting humans worldwide. Timely detection can assist in the reduction of both mortality rates and treatment costs. Segmentation plays a crucial role in the timely detection of breast cancer. This study describes the methodology and findings of a systematic review (SR) that aimed to evaluate cutting-edge segmentation approaches in computer-aided diagnosis (CAD). The primary objective is to provide a general overview and analysis of each of these methods, as well as to address the limitations and difficulties associated with them. Based on the observed shortcomings, directions for future investigation have also been offered. In addition, this SR also provides an overview of numerous publicly available benchmark datasets used in the reviewed literature. Unsupervised algorithms have been found to be more commonly used than supervised and DL counterparts. DL techniques require powerful computers and enormous volumes of training data. In most of the existing segmentation approaches, digitised images from DDSM and mini-MIAS datasets have been employed for experimentation. Both medical professionals and researchers who are aiming to address research gaps and make significant contributions to breast cancer diagnoses may find this systematic review to be valuable.

Dynamic Network-Assisted D2D-Aided Coded Distributed Learning

Today, numerous machine learning (ML) applications offer continuous data processing and real-time data analytics at the edge of wireless networks. Distributed real-time ML solutions are highly susceptible to the so-called straggler effect caused by resource heterogeneity, which can be mitigated by various computation offloading mechanisms that severely impact communication efficiency, especially in large-scale scenarios. To reduce the communication overhead, we leverage device-to-device (D2D) connectivity, which enhances spectrum utilization and allows for efficient data exchange between proximate devices. In particular, we design a novel D2D-aided coded distributed learning method named D2D-CDL for efficient load balancing across devices. The proposed solution captures system dynamics, including <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">data</i> (time-varying learning model, irregular intensity of data arrivals), <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">device</i> (diverse computational resources and volume of training data), and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">deployment</i> (different locations and D2D graph connectivity). To decrease the number of communication rounds, we derive an optimal compression rate, which minimizes the processing time. The resulting optimization problem provides suboptimal compression parameters that improve the total training time. Our proposed method is particularly beneficial for real-time collaborative applications, where users continuously generate training data thus yielding a model drift.

A Physics-Constrained Bayesian neural network for battery remaining useful life prediction

In order to predict the remaining useful life (RUL) of lithium-ion batteries, a capacity degradation model may be developed using either simplified physical laws or machine learning-based methods. It is observed that even though degradation models based on simplified physical laws are easy to implement, they may result in large error in the application of failure prognostics. While data-driven prognostics models can provide more accurate degradation forecasting, they may require a large volume of training data and may invoke predictions inconsistent with physical laws. It is also very challenging for existing methods to predict the RUL at the early stages of battery life. In this paper, we propose a Bayesian physics-constrained neural network for battery RUL prediction by overcoming limitations of the current methods. In the proposed method, a neural differential operator is learned from the first 100 cycles of data. The neural differential operator is modeled with a Bayesian neural network architecture that separates the fixed history dependence from the time dependence to isolate epistemic uncertainty quantification. Using the battery dataset presented in the paper by Severson et al. as an example, we compare our proposed method with a simplified physics-based degradation forecasting model and two data-driven prognostics models. The results show that the proposed physics-constrained neural network can provide more accurate RUL estimation than the other methods with the same group of training data. Most importantly, the proposed method allows for RUL prediction at earlier stages of the battery life cycle.

Generating synthetic samples of chondrosarcoma histopathology with a denoising diffusion probabilistic model.

e13592 Background: The emergence of digital pathology — an image-based environment for the acquisition, management and interpretation of pathology information supported by computational techniques for data extraction and analysis — is changing the pathology ecosystem. The development of machine-learning approaches for the extraction of information from image data, allows for tissue interrogation in a way that was not previously possible. However, creating digital pathology algorithms requires large volumes of training data, often on the order of thousands of histopathology slides. This becomes problematic for rare diseases, where imaging datasets of such size do not exist. This makes it impossible to train digital pathology models for these rare conditions. However, recent advances in generative deep learning models may provide a method for overcoming this lack of histology data for rare diseases. Pre-trained diffusion-based probabilistic models can be used to create photorealistic variations of existing images. In this study, we explored the potential of using a deep generative model created by OpenAI for the purpose of producing synthetic histopathology images, using chondrosarcoma as our rare tumor of interest. Methods: Our team compiled a dataset of 55 chondrosarcoma histolopathology images from the annotated records of Dr. Henry Jaffe, a pioneering authority in musculoskeletal pathology. We built a deep learning image-generation application in a Jupyter notebook environment, iterating upon OpenAI’s DALL-E application processing interface (API) with python programming language. Using the chondrosarcoma histology dataset and NVIDIA GPUs, we trained the deep learning application to generate multiple synthetic variations of each real chondrosarcoma image. Results: After several hours, the deep learning model successfully generated 1,000 images of chondrosarcoma from 55 original images. The synthetic histology images retained photorealistic quality and displayed characteristic cellular features of chondrosarcoma tumor tissue. Conclusions: Deep generative models may be useful in addressing issues of data scarcity in rare diseases, such as chondrosarcoma. For example, in situations where existing imaging data is insufficient for training diagnostic computer vision models, diffusion-based generative models could be applied to create training datasets. However, further exploration of ethical considerations and qualitative analyses of these generated data are needed.

Interpretable Machine Learning Models for Phase Prediction in Polymerization-Induced Self-Assembly.

While polymerization-induced self-assembly (PISA) has become a preferred synthetic route toward amphiphilic block copolymer self-assemblies, predicting their phase behavior from experimental design is extremely challenging, requiring time and work-intensive creation of empirical phase diagrams whenever self-assemblies of novel monomer pairs are sought for specific applications. To alleviate this burden, we develop here the first framework for a data-driven methodology for the probabilistic modeling of PISA morphologies based on a selection and suitable adaption of statistical machine learning methods. As the complexity of PISA precludes generating large volumes of training data with in silico simulations, we focus on interpretable low variance methods that can be interrogated for conformity with chemical intuition and that promise to work well with only 592 training data points which we curated from the PISA literature. We found that among the evaluated linear models, generalized additive models, and rule and tree ensembles, all but the linear models show a decent interpolation performance with around 0.2 estimated error rate and 1 bit expected cross entropy loss (surprisal) when predicting the mixture of morphologies formed from monomer pairs already encountered in the training data. When considering extrapolation to new monomer combinations, the model performance is weaker but the best model (random forest) still achieves highly nontrivial prediction performance (0.27 error rate, 1.6 bit surprisal), which renders it a good candidate to support the creation of empirical phase diagrams for new monomers and conditions. Indeed, we find in three case studies that, when used to actively learn phase diagrams, the model is able to select a smart set of experiments that lead to satisfactory phase diagrams after observing only relatively few data points (5-16) for the targeted conditions. The data set as well as all model training and evaluation codes are publicly available through the GitHub repository of the last author.

Combined visual and spatial-temporal information for appearance change person re-identification

Person re-identification (ReID) seeks to identify the same individual across different cameras by matching their corresponding images. The current ReID datasets are limited in size and diversity, especially in terms of clothing changes, making traditional techniques vulnerable to appearance variations. Further, current approaches rely heavily on appearance features for discrimination, which is unreliable when a person’s appearance changes. We hypothesize that the ReID accuracy can be enhanced by training the ReID model on a large volume of diversified training data and combining multiple features for discrimination. We use the image channel shuffling data augmentation method, producing a large volume of diversified training data. Also, a two-stream visual and spatial-temporal method is proposed to learn the feasible features for appearance change scenarios. The appearance features obtained from the visual stream are combined with spatio-temporal information to discriminate between two people. The proposed approach is evaluated for its robustness on short-term and on long-term datasets. The presented two-stream approach outperforms earlier methods by achieving Rank1 accuracy of 98.6% on Market1501, 95.52% on DukeMTMC-reID, 76.21% on LTCC, and 91.76% on VC-Clothes, respectively.

Two-Module Neural Network Method of Information Processing in Gas Analysis Systems

Purpose of research: reduction of additional errors in measuring gas concentrations in gas analytical systems (GS) caused by the sensitivity of semiconductor sensors to non-target components of gas mixtures, ambient temperature and humidity. To develop and test a two-module neural network method for processing information in a GS, which allows automating the processes of generating training data and searching for the optimal structure of artificial neural networks (ANNs), reducing errors in reproducing the characteristics of sensors by replacing their mathematical models with neural networks.Methods. Theory of artificial neural networks, numerical methods, simulation methods. To evaluate the effectiveness of the proposed solution, the relative error (d), standard deviation (RMS) were calculated, and comparison with analogues was carried out.Results: a two-module neural network method for processing information in a GS has been studied. Numerical modeling was used to carry out experimental studies on the choice of optimal ANN structures, the volume and composition of training data. In the course of experimental studies, the errors of generating training data using ANN (less than 5%) and determining the concentrations of detected gases under conditions of fluctuations in the parameters of the air environment and the composition of the gas mixture (less than 4%) were calculated.Conclusion. A two-module neural network method for information processing is proposed, which is distinguished by the use of two successive modules of multilayer neural networks for generating training data and processing information coming from the GS sensor unit. The use of an auxiliary module makes it possible to compress the initial data, unify and automate the process of their generation, as well as improve the accuracy of reproduction of multiparameter sensor conversion functions, in comparison with alternative methods. Results of experimental studies of the effectiveness of using the information processing method to reduce additional errors in the quantitative determination of the composition of the air environment under conditions of parameter fluctuations are presented.

CSI-Based Human Activity Recognition Using Multi-Input Multi-Output Autoencoder and Fine-Tuning.

Wi-Fi-based human activity recognition (HAR) has gained considerable attention recently due to its ease of use and the availability of its infrastructures and sensors. Channel state information (CSI) captures how Wi-Fi signals are transmitted through the environment. Using channel state information of the received signals transmitted from Wi-Fi access points, human activity can be recognized with more accuracy compared with the received signal strength indicator (RSSI). However, in many scenarios and applications, there is a serious limit in the volume of training data because of cost, time, or resource constraints. In this study, multiple deep learning models have been trained for HAR to achieve an acceptable accuracy level while using less training data compared to other machine learning techniques. To do so, a pretrained encoder which is trained using only a limited number of data samples, is utilized for feature extraction. Then, by using fine-tuning, this encoder is utilized in the classifier, which is trained by a fraction of the rest of the data, and the training is continued alongside the rest of the classifier's layers. Simulation results show that by using only 50% of the training data, there is a 20% improvement compared with the case where the encoder is not used. We also showed that by using an untrainable encoder, an accuracy improvement of 11% using 50% of the training data is achievable with a lower complexity level.

Efficient Data-Driven Off-Design Constraint Modeling for Practical Aerodynamic Shape Optimization

Off-design constraints are essential in practical aerodynamic shape optimization. Physics-based data-driven modeling has shown to be a feasible way to formulate generalizable off-design constraints. However, two issues hinder the adoption of this approach: inadequate physical mechanism studies and the requirement of massive training data. To address these issues, we propose a feature-oriented data-driven model to learn fundamental physical mechanisms from high-dimensional data. This is achieved by finding low-dimensional latent space relevant to the investigated off-design performance. Then, we customize the constraint model based on the learned physical mechanisms, improving generalizability without relying on prior knowledge. We also propose a Bayesian-optimization-based sampling method to adjust the training data distribution, prioritizing samples with good aerodynamic performance. This uneven sampling strategy improves data efficiency by ensuring accuracy when approaching optimal aerodynamic shapes. The effectiveness of the proposed methods is shown in a low-Reynolds-number airfoil design optimization case and a transonic airfoil design optimization case. We obtain generalizable data-driven off-design aerodynamic models with no prior physical studies, and we reduce the training data volume by 95% compared with a conventional data-driven approach. This work lays the technical foundation for sample-efficient and generalizable data-driven modeling of off-design aerodynamic constraints.

Research on group distribution leakage analysis method for RSM

Side channel attack is a new and simple attack idea, and its proposal has created a significant threat to the security of cryptographic devices. With the continuous development of deep learning techniques, deep learning-based side channel attacks have enabled the attack techniques to reach new heights, but the attack efficiency and training time have become new problems. In order to improve the attack efficiency, researchers have proposed a large number of leakage analysis methods, which can effectively extract the intervals with significant leakage in the energy consumption curve and reduce the training parameters and data volume. In order to deal with side channel attacks, researchers have proposed various resistance strategies, among which masking protection is the simplest and widely used way. RSM rotating S-box masking scheme is a particularly important masking protection scheme, which uses a first-order masking scheme to randomize the leakage energy consumption of sensitive information using random masks, thus making the algorithm resistant to first-order side channel attacks. For this type of masking algorithm, this paper proposes a new leakage analysis algorithm, namely "Group Distribution Difference" (GDD). This algorithm is based on the distribution difference of energy consumption, and the energy consumption frequency in the group is calculated instead of the probability distribution. The KL distance is used to calculate the difference between groups and find out the leakage interval, so as to achieve an efficient attack.

Brain-machine interface based on transfer-learning for detecting the appearance of obstacles during exoskeleton-assisted walking.

Brain-machine interfaces (BMIs) attempt to establish communication between the user and the device to be controlled. BMIs have great challenges to face in order to design a robust control in the real field of application. The artifacts, high volume of training data, and non-stationarity of the signal of EEG-based interfaces are challenges that classical processing techniques do not solve, showing certain shortcomings in the real-time domain. Recent advances in deep-learning techniques open a window of opportunity to solve some of these problems. In this work, an interface able to detect the evoked potential that occurs when a person intends to stop due to the appearance of an unexpected obstacle has been developed. First, the interface was tested on a treadmill with five subjects, in which the user stopped when an obstacle appeared (simulated by a laser). The analysis is based on two consecutive convolutional networks: the first one to discern the intention to stop against normal walking and the second one to correct false detections of the previous one. The results were superior when using the methodology of the two consecutive networks vs. only the first one in a cross-validation pseudo-online analysis. The false positives per min (FP/min) decreased from 31.8 to 3.9 FP/min and the number of repetitions in which there were no false positives and true positives (TP) improved from 34.9% to 60.3% NOFP/TP. This methodology was tested in a closed-loop experiment with an exoskeleton, in which the brain-machine interface (BMI) detected an obstacle and sent the command to the exoskeleton to stop. This methodology was tested with three healthy subjects, and the online results were 3.8 FP/min and 49.3% NOFP/TP. To make this model feasible for non-able bodied patients with a reduced and manageable time frame, transfer-learning techniques were applied and validated in the previous tests, and were then applied to patients. The results for two incomplete Spinal Cord Injury (iSCI) patients were 37.9% NOFP/TP and 7.7 FP/min.

Using Geospatial Information to Map Yield Gain from the Use of Azospirillum brasilense in Furrow

The application of biological products in agricultural crops has become increasingly prominent. The growth-promoting bacterium Azospirillum brasilense has been used as an alternative to promote greater yield in maize crops. In the context of precision agriculture, interpreting geospatial data has allowed for monitoring the effect of the application of products that increase the yield of corn crops. The objective of this work was to evaluate the potential of Kriging techniques and spectral models through images in estimating the gain in yield of maize crop after applying A. brasilense. Analyses were carried out in two commercial areas treated with A. brasilense. The results revealed that models of yield prediction by Kriging with a high volume of training data estimated the yield gain with a root-mean-square error deviation (RMSE%), mean absolute percentage error (MAPE%), and R2 to be 6.67, 5.42, and 0.88, respectively. For spectral models with a low volume of training data, yield gain was estimated with RMSE%, MAPE%, and R2 to be 9.3, 7.71, and 0.80, respectively. The results demonstrate the potential to map the spatial distribution of productivity gains in corn crops following the application of A. brasilense.

Transferability investigation of a Sim2Real deep transfer learning framework for cross-building energy prediction

Data-driven methods generally require large-scale historical data volume to obtain a better building energy prediction performance. Transfer learning methods have been widely used to address this issue by utilizing knowledge learned from similar tasks. However, most studies on building energy prediction are mainly focused on the effect of a single transfer learning model. There are few studies to systematically investigate the effect of building types, climate zones and data volume on energy prediction performance by using different transfer learning models with the auxiliary simulation data. This study proposes a Sim2Real transfer learning framework using simulation datasets for building energy prediction to fill this research gap. Three different experiment scenarios are conducted to investigate the effect of different building types, climate zones and data volume on the prediction performance from the perspective of simulation to reality by using the proposed Sim2Real transfer learning framework. Extensive experiment cases demonstrate that the proposed Sim2Real transfer learning framework can effectively enhance the MAPE by 1.06%∼29.05% with proper transfer learning models compared with the baseline model LSTM. In addition, transfer learning strategies are not necessary when the training data volume of target building is over five weeks. The proposed Sim2Real transfer learning framework can provide the bridge to enhance the building energy prediction performance from the perspective of simulation to reality.

Blinded Predictions and Post Hoc Analysis of the Second Solubility Challenge Data: Exploring Training Data and Feature Set Selection for Machine and Deep Learning Models.

Accurate methods to predict solubility from molecular structure are highly sought after in the chemical sciences. To assess the state of the art, the American Chemical Society organized a "Second Solubility Challenge" in 2019, in which competitors were invited to submit blinded predictions of the solubilities of 132 drug-like molecules. In the first part of this article, we describe the development of two models that were submitted to the Blind Challenge in 2019 but which have not previously been reported. These models were based on computationally inexpensive molecular descriptors and traditional machine learning algorithms and were trained on a relatively small data set of 300 molecules. In the second part of the article, to test the hypothesis that predictions would improve with more advanced algorithms and higher volumes of training data, we compare these original predictions with those made after the deadline using deep learning models trained on larger solubility data sets consisting of 2999 and 5697 molecules. The results show that there are several algorithms that are able to obtain near state-of-the-art performance on the solubility challenge data sets, with the best model, a graph convolutional neural network, resulting in an RMSE of 0.86 log units. Critical analysis of the models reveals systematic differences between the performance of models using certain feature sets and training data sets. The results suggest that careful selection of high quality training data from relevant regions of chemical space is critical for prediction accuracy but that other methodological issues remain problematic for machine learning solubility models, such as the difficulty in modeling complex chemical spaces from sparse training data sets.

Classification of Complicated Urban Forest Acoustic Scenes with Deep Learning Models

The use of passive acoustic monitoring (PAM) can compensate for the shortcomings of traditional survey methods on spatial and temporal scales and achieve all-weather and wide-scale assessment and prediction of environmental dynamics. Assessing the impact of human activities on biodiversity by analyzing the characteristics of acoustic scenes in the environment is a frontier hotspot in urban forestry. However, with the accumulation of monitoring data, the selection and parameter setting of the deep learning model greatly affect the content and efficiency of sound scene classification. This study compared and evaluated the performance of different deep learning models for acoustic scene classification based on the recorded sound data from Guangzhou urban forest. There are seven categories of acoustic scenes for classification: human sound, insect sound, bird sound, bird–human sound, insect–human sound, bird–insect sound, and silence. A dataset containing seven acoustic scenes was constructed, with 1000 samples for each scene. The requirements of the deep learning models on the training data volume and training epochs in the acoustic scene classification were evaluated through several sets of comparison experiments, and it was found that the models were able to achieve satisfactory accuracy when the training sample data volume for a single category was 600 and the training epochs were 100. To evaluate the generalization performance of different models to new data, a small test dataset was constructed, and multiple trained models were used to make predictions on the test dataset. All experimental results showed that the DenseNet_BC_34 model performs best among the comparison models, with an overall accuracy of 93.81% for the seven acoustic scenes on the validation dataset. This study provides practical experience for the application of deep learning techniques in urban sound monitoring and provides new perspectives and technical support for further exploring the relationship between human activities and biodiversity.

Design De-Identification of Thermal History for Collaborative Process-Defect Modeling of Directed Energy Deposition Processes

Abstract There is an urgent need for developing collaborative process-defect modeling in metal-based additive manufacturing (AM). This mainly stems from the high volume of training data needed to develop reliable machine learning models for in-situ anomaly detection. The requirements for large data are especially challenging for small-to-medium manufacturers (SMMs), for whom collecting copious amounts of data is usually cost prohibitive. The objective of this research is to develop a secured data sharing mechanism for directed energy deposition (DED) based AM without disclosing product design information, facilitating secured data aggregation for collaborative modeling. However, one major obstacle is the privacy concerns that arise from data sharing, since AM process data contain confidential design information, such as the printing path. The proposed adaptive design de-identification for additive manufacturing (ADDAM) methodology integrates AM process knowledge into an adaptive de-identification procedure to mask the printing trajectory information in metal-based AM thermal history, which otherwise discloses substantial printing path information. This adaptive approach applies a flexible data privacy level to each thermal image based on its similarity with the other images, facilitating better data utility preservation while protecting data privacy. A real-world case study was used to validate the proposed method based on the fabrication of two cylindrical parts using a DED process. These results are expressed as a Pareto optimal solution, demonstrating significant improvements in privacy gain and minimal utility loss. The proposed method can facilitate privacy improvements of up to 30% with as little as 0% losses in dataset utility after de-identification.

Developing efficient deep learning model for predicting copolymer properties.

Deep learning models are gaining popularity and potency in predicting polymer properties. These models can be built using pre-existing data and are useful for the rapid prediction of polymer properties. However, the performance of a deep learning model is intricately connected to its topology and the volume of training data. There is no facile protocol available to select a deep learning architecture, and there is a lack of a large volume of homogeneous sequence-property data of polymers. These two factors are the primary bottleneck for the efficient development of deep learning models for polymers. Here we assess the severity of these factors and propose strategies to address them. We show that a linear layer-by-layer expansion of a neural network can help in identifying the best neural network topology for a given problem. Moreover, we map the discrete sequence space of a polymer to a continuous one-dimensional latent space using a feature extraction technique to identify minimal data points for training a deep learning model. We implement these approaches for two representative cases of building sequence-property surrogate models, viz., the single-molecule radius of gyration of a copolymer and copolymer compatibilizer. This work demonstrates efficient methods for building deep learning models with minimal data and hyperparameters for predicting sequence-defined properties of polymers.

Data-Driven Optimal Power Dispatch for Distributed Energy Resources in Radial Feeder using Multi-Stage Regression

This paper develops a multi-stage regression (MSR) based algorithm for dispatching distributed energy resources (DERs) in a power distribution system. Application of the standard regression algorithm to network power flow problems generally suffer from performance limitations due to collinearity in training data, (i.e. voltages, currents and nodal demands), and non-linearity inherent in power flows and DER dispatches. Here, regression collinearity is addressed with intelligent choice of the input data training set which also considerably reduces the requirements on the volume of training data. Next, a logistic regression based labeling scheme is developed and applied to classify the data into disjoint training sets which significantly improves the prediction accuracy. Simulations are provided to demonstrate the performance of the proposed algorithm in a radial distribution system voltage support application under a range of operating conditions and considering uncertainty in parameter values. The proposed approach and learnings can be extended to a range of network power flow problems and DER-based applications.

Development of domain-specific automatic speech recognition models based on open-source data

Currently, deep neural network architectures are used in the development of automatic speech recognition models for specialized subject areas, which require a large volume of training data. However, these models often prove to be poorly suitable for use in specific information systems due to the inadequate recognition of specialized subject vocabularies. Further training of models to improve their quality in a specific context of recognition encounters difficulties in obtaining sufficient data and the laboriousness of their labeling. Given this, a pertinent task is the creation of methods that allow for a reduction in the effort required to build applied speech recognition models and improve their quality when used in various subject areas. The study used language model-based text topic modeling techniques to adapt open-source data. A deep neural network was used as the pre-trained speech recognition model. For training, data sets from open sources were utilized. A method has been developed for creating automatic speech recognition models for specialized subject areas. This method involves an intermediate step of learning the lexicon of the subject area from data obtained from open sources, selected using thematic sampling. Based on this method, automatic speech recognition models for the energy and healthcare sectors were created and researched, demonstrating higher recognition results than models developed through traditional methods. The validation of the proposed method confirmed its effectiveness. The applied neural network models, developed based on this method, demonstrated their ability to operate in information systems of energy and healthcare facilities in Russian, English, and Italian languages without additional training on closed data.