bsuirДоклады БГУИРDoklady BGUIR1729-76482708-0382БГУИР10.35596/1729-7648-2026-24-4-81-88bsuir-4407Research ArticleСтатьиВыбор архитектуры нейронной сети в задаче классификацииNeural Network Architecture Selection in Classification ProblemМацкевичВ. В.MatskevichV.

Мацкевич Вадим Владимирович, канд. техн. наук, доц. каф. информационных систем управления

220030, Минск, просп. Независимости, 4

Тел.: +375 29 125-49-07

Matskevich Vadim, Cand. Sci. (Tech.), Associate Professor of the Department of Information Management Systems

220030, Minsk, Nezavisimosti Ave., 4

matskevich1997@gmail.comБелорусский государственный университетБеларусьBelarusian State UniversityBelarus2026300820262448188Copyright © Мацкевич В.В., 20262026Мацкевич В.В.Matskevich V.Данная работа распространяется под лицензией Creative Commons Attribution 4.0.This work is licensed under a Creative Commons Attribution 4.0 License.https://doklady.bsuir.by/jour/article/view/4407

Нейросетевая обработка данных большой размерности требует значительных вычислительных ресурсов. Неверный выбор архитектуры нейронной сети или преобразования данных может привести к существенным потерям времени и ресурсов на повторное обучение сети. В статье предлагается алгоритм выбора архитектуры сжимающих слоев глубокой доверительной сети и преобразования данных на основе анализа многомерной матрицы несоответствий. Введенный обобщенный коэффициент перекоса описывает равномерность распределения объектов между классами нейронной сети и позволяет численно оценить эффективность архитектуры сжимающих слоев сети и преобразования данных. На примере прикладной задачи показано, что предложенный алгоритм повышает качество ее решения более чем на 10 % путем выбора наиболее эффективной архитектуры нейросети.

Neural network processing of high-dimensional data requires significant computing resources. An incorrect choice of neural network architecture or data transformation can lead to significant losses of time and resources in network retraining. This article proposes an algorithm for selecting the architecture of compressive layers for a deep belief network and data transformation based on the analysis of a multi-label confusion matrix. The introduced generalized skewness coefficient describes the uniformity of the distribution of objects between neural network classes and allows for a numerical evaluation of the effectiveness of the network’s compressive layer architecture and data transformation. Using an applied problem as an example, it is demonstrated that the proposed algorithm improves the quality of its solution by more than 10 % by selecting the most effective neural network architecture.

входные данныенейронные сетиобучениеархитектурапреобразование данныхмногомерная матрица несоответствийобобщенный коэффициент перекосаinput dataneural networkstrainingarchitecturedata transformationmulti-label confusion matrixgeneralized skewness coefficientReferencesHeydarian, M. MLCM: Multi-Label Confusion Matrix / M. Heydarian, T. E. Doyle, R. Samavi // IEEE Access. 2022. Vol. 10. P. 19083–19095. http://dx.doi.org/10.1109/ACCESS.2022.3151048.Heydarian M., Doyle T. E., Samavi R. (2022) MLCM: Multi-Label Confusion Matrix. IEEE Access. 10, 19083–19095. http://dx.doi.org/10.1109/ACCESS.2022.3151048.Мацкевич, В. В. Адаптация нейросетевого классификатора к неустойчивым входным данным / В. В. Мацкевич, Ц. Го // Доклады БГУИР. 2025. Т. 23, № 5. С. 99–104. http://dx.doi.org/10.35596/1729-7648-2025-23-5-99-104.Matskevich V. V., Guo J. (2025) Neural Network Classifier Adaptation to Unstable Input Data. Doklady BGUIR. 23 (5), 99–104. http://dx.doi.org/10.35596/1729-7648-2025-23-5-99-104 (in Russian).AutoPolCNN: A Neural Architecture Search Method of Convolutional Neural Network for Polsar Image Classification / G. Liu [et al.] // Knowledge-Based Systems. 2025. Vol. 312. https://doi.org/10.1016/j.knosys.2025.113122.Liu G., Li Y., Chen Y., Shang R., Jiao L. (2025) AutoPolCNN: A Neural Architecture Search Method of Convolutional Neural Network for PolSAR Image Classification. Knowledge-Based Systems. 312. https:// doi.org/10.1016/j.knosys.2025.113122.Automatic Neural Network Construction Based on Neural Tangent Kernel for IRS-Aided Beamforming / H. Shi [et al.] // IEEE Transactions on Wireless Communications. 2025. Vol. 25. P. 1155–1168. https://doi.org/10.1109/TWC.2025.3589367.Shi H., Ji T., Wang Z., Jin S., Huang Y. (2025) Automatic Neural Network Construction Based on Neural Tangent Kernel for IRS-Aided Beamforming. IEEE Transactions on Wireless Communications. 25, 1155–1168. https://doi.org/10.1109/TWC.2025.3589367.Synthesis of Convolutional Neural Network Architectures for Biomedical Image Classification / O. Berezsky [et al.] // Biomedical Signal Processing and Control. 2024. Vol. 95, Part B. https://doi.org/10.1016/j.bspc.2024.106325.Berezsky O., Liashchynskyi P., Pitsun O., Izonin I. (2024) Synthesis of Convolutional Neural Network Architectures for Biomedical Image Classification. Biomedical Signal Processing and Control. 95 (B). https:// doi.org/10.1016/j.bspc.2024.106325.Akay, B. A Comprehensive Survey on Optimizing Deep Learning Models by Metaheuristics / B. Akay, D. Karaboga, R. Akay // Artificial Intelligence Review. 2022. Vol. 55, Iss. 2. https://doi.org/10.1007/s10462-021-09992-0.Akay B., Karaboga D., Akay R. (2022) A Comprehensive Survey on Optimizing Deep Learning Models by Metaheuristics. Artificial Intelligence Review. 55 (2). https://doi.org/10.1007/s10462-021-09992-0.Xi, Sh. Dynamic Gaussian Deep Belief Network Design and Stock Market Application / Sh. Xi, X. Xu // Intelligent Data Analysis. 2023. Vol. 27, Iss. 2. P. 519–534. http://dx.doi.org/10.3233/ida-216340.Xi Sh., Xu X. (2023) Dynamic Gaussian Deep Belief Network Design and Stock Market Application. Intelligent Data Analysis. 27 (2), 519–534. http://dx.doi.org/10.3233/ida-216340.Zhou, X. Landscape’s Non-Natural Changes Detection System by Satellites Images Based on Local Areas / X. Zhou [et al.] // Pattern Recognition and Image Analysis. Advances in Mathematical Theory and Applications. 2024. Vol. 34, Iss. 2. P. 365–378, https://doi.org/10.1134/S1054661824700159.Zhou X., Bu Q., Matskevich V., Nedzved A. (2024) Landscape’s Non-Natural Changes Detection System by Satellites Images Based on Local Areas. Pattern Recognition and Image Analysis. 34 (2), 365–378. http:// dx.doi.org/10.1134/S1054661824700159.Krasnoproshin, V. V. Random Search in Neural Networks Training / V. V. Krasnoproshin, V. V. Matskevich // Pattern Recognition and Image Analysis. 2024. Vol. 34, Iss. 2. P. 309–316. https://doi.org/10.1134/S105466182470010X.Krasnoproshin V. V., Matskevich V. V. (2024) Random Search in Neural Networks Training. Pattern Recognition and Image Analysis. 34 (2), 309–316. https://doi.org/10.1134/S105466182470010X.Tang, H. A Pure MLP-Mixer-Based GAN Framework for Guided Image Translation / H. Tang, B. Ren, N. Sebe // Pattern Recognition. 2025. Vol. 157. https://doi.org/10.1016/j.patcog.2024.110894.Tang H., Ren B., Sebe N. (2025) A Pure MLP-Mixer-Based GAN Framework for Guided Image Translation. Pattern Recognition. 157. https://doi.org/10.1016/j.patcog.2024.110894.Multi-Scale MLP-Mixer for Image Classification. / H. Zhang [et al.] // Knowledge-Based Systems. 2022. Vol. 258. https://doi.org/10.1016/j.knosys.2022.109792.Zhang H., Dong Z., Li B., He S. (2022) Multi-Scale MLP-Mixer for Image Classification. Knowledge-Based Systems. 258. https://doi.org/10.1016/j.knosys.2022.109792.Raj, R. An Extensive Study of Convolutional Neural Networks: Applications in Computer Vision for Improved Robotics Perceptions / R. Raj, A. Kos // Sensors. 2025. Vol. 25, Iss. 4. https://doi.org/10.3390/s25041033.Raj R., Kos A. (2025) An Extensive Study of Convolutional Neural Networks: Applications in Computer Vision for Improved Robotics Perceptions. Sensors. 25 (4). https://doi.org/10.3390/s25041033.Abdellatef, E. Detection of Human Activities Using Multi-Layer Convolutional Neural Network / E. Abdellatef, R. M. Al-Makhlasawy, W. A. Shalaby // Scientific Reports. 2025. Vol. 15, Iss. 1. https://doi.org/10.1038/s41598-025-90307-6.Abdellatef E., Al-Makhlasawy R. M., Shalaby W. A. (2025) Detection of Human Activities Using Multi-Layer Convolutional Neural Network. Scientific Reports. 15 (1). https://doi.org/10.1038/s41598-025-90307-6.Lee, M. GELU Activation Function in Deep Learning: A Comprehensive Mathematical Analysis and Performance / M. Lee // arXiv. 2023. https://doi.org/10.48550/arXiv.2305.12073.Lee M. (2023) GELU Activation Function in Deep Learning: A Comprehensive Mathematical Analysis and Performance. arXiv. https://doi.org/10.48550/arXiv.2305.12073.

The authors declare that there are no conflicts of interest present.