Cited by Improved Architecture and the Synthesis Algorithm for Bithreshold Neural Network Classifier

Transfer Learning Approach for Human Activity Recognition Based on Continuous Wavelet Transform DOI

Olena Pavliuk, Myroslav Mishchuk, Christine Strauß

et al.

Algorithms, Journal Year: 2023, Volume and Issue: 16(2), P. 77 - 77

Published: Feb. 1, 2023

Over the last few years, human activity recognition (HAR) has drawn increasing interest from scientific community. This attention is mainly attributable to proliferation of wearable sensors and expanding role HAR in such fields as healthcare, sports, monitoring. Convolutional neural networks (CNN) are becoming a popular approach for addressing problems. However, this method requires extensive training datasets perform adequately on new data. paper proposes novel deep learning model pre-trained scalograms generated using continuous wavelet transform (CWT). Nine CNN architectures different CWT configurations were considered select best performing combination, resulting evaluation more than 300 models. On source KU-HAR dataset, selected achieved classification accuracy an F1 score 97.48% 97.52%, respectively, which outperformed contemporary state-of-the-art works where dataset was employed. target UCI-HAPT proposed resulted maximum F1-score increase 0.21% 0.33%, whole 2.82% 2.89%, subset. It concluded that usage model, particularly with frozen layers, results improved performance, faster training, smoother gradient descent small datasets. use sufficiently large may lead negative transfer degradation.

Language: Английский

Citations

A cascade ensemble-learning model for the deployment at the edge: case on missing IoT data recovery in environmental monitoring systems DOI

Ivan Izonin, Roman Tkachenko, Юрий Крак

et al.

Frontiers in Environmental Science, Journal Year: 2023, Volume and Issue: 11

Published: Oct. 26, 2023

In recent years, more and applied industries have relied on data collection by IoT devices. Various devices generate vast volumes of that require efficient processing. Usually, the intellectual analysis such takes place in centers cloud environments. However, problems transferring large long wait for a response from center further corrective actions system led to search new processing methods. One possible option is Edge computing. Intelligent places their eliminates disadvantages mentioned above, revealing many advantages using an approach practice. computing challenging implement when different collect independent attributes required classification/regression. order overcome this limitation, authors developed cascade ensemble-learning model deployment at Edge. It based principles cascading machine learning methods, where each device collects performs its it contains. The results work are transmitted next device, which analyzes collects, taking into account output previous device. All at-tributes taken way. Because this, proposed provides: 1) possibility effective implementation intelligent analysis, is, even before transmission center; 2) increasing, some cases maintaining, classification/regression accuracy same level can be achieved 3) significantly reducing duration training procedures due smaller number simulation was performed real-world set data. missing recovery task atmospheric air state solved. selected optimal parameters approach. established provides slight increase prediction while procedure. case, main advantage all happens within bounds computing, opens up several benefits

Language: Английский

Citations

Optimizing Neural Networks for Chemical Reaction Prediction: Insights from Methylene Blue Reduction Reactions DOI

Ivan Malashin, В С Тынченко, Andrei Gantimurov

et al.

International Journal of Molecular Sciences, Journal Year: 2024, Volume and Issue: 25(7), P. 3860 - 3860

Published: March 29, 2024

This paper offers a thorough investigation of hyperparameter tuning for neural network architectures using datasets encompassing various combinations Methylene Blue (MB) Reduction by Ascorbic Acid (AA) reactions with different solvents and concentrations. The aim is to predict coefficients decay plots MB absorbance, shedding light on the complex dynamics chemical reactions. Our findings reveal that optimal model, determined through our investigation, consists five hidden layers, each sixteen neurons employing Swish activation function. model yields an NMSE 0.05, 0.03, 0.04 predicting A, B, C, respectively, in exponential equation A + B · e−x/C. These contribute realm drug design based machine learning, providing valuable insights into optimizing reaction predictions.

Language: Английский

Citations

A non-linear SVR-based cascade model for improving prediction accuracy of biomedical data analysis DOI

Ivan Izonin, Roman Tkachenko, Oleksandr Gurbych

et al.

Mathematical Biosciences & Engineering, Journal Year: 2023, Volume and Issue: 20(7), P. 13398 - 13414

Published: Jan. 1, 2023

<abstract> <p>Biomedical data analysis is essential in current diagnosis, treatment, and patient condition monitoring. The large volumes of that characterize this area require simple but accurate fast methods intellectual to improve the level medical services. Existing machine learning (ML) many resources (time, memory, energy) when processing datasets. Or they demonstrate a accuracy insufficient for solving specific application task. In paper, we developed new ensemble model increased approximation problems biomedical sets. based on cascading ML response surface linearization principles. addition, used Ito decomposition as means nonlinearly expanding inputs at each model. As weak learners, Support Vector Regression (SVR) with linear kernel was due significant advantages demonstrated by method among existing ones. training procedures SVR-based cascade are described, flow chart its implementation presented. modeling carried out real-world tabular set volume. task predicting heart rate individuals solved, which provides possibility determining human stress, an indicator various applied fields. optimal parameters operating were selected experimentally. authors shown more than 20 times higher (according Mean Squared Error (MSE)), well reduction duration procedure compared method, provided highest work those considered.</p> </abstract>

Language: Английский

Citations

Multithreshold Neural Units and Networks DOI

Vladyslav Kotsovsky, Anatoliy Batyuk

Published: Oct. 19, 2023

We deal with theoretical issues concerning the application of multithreshold architecture in theory neural computation. The way representing a function by 2-layer network consisting single-threshold units equal weights is established paper. also study complexity problem learning k-threshold neurons and prove that this NP-hard if number thresholds greater than one.

Language: Английский

Citations

Improved Architecture and the Synthesis Algorithm for Bithreshold Neural Network Classifier DOI

Vladyslav Kotsovsky, Anatoliy Batyuk

Published: Oct. 19, 2023

The model of the 3-layer feed-forward neural network is introduced whose first hidden layer consists bithreshold neurons and other layers—of single-threshold ones. proposed capable to recognize compact finite set patterns using a union hyperrectangular decision regions in n-dimensional space. We design multiclass classifier on base such network, propose synthesis algorithm for it estimate networks size as well time computations. simulation results demonstrate that application additional improves accuracy classification.

Language: Английский

Citations