S.S. Svyatkin1, A.K. Vishnevsky2, I.K. Novikov3, I.V. Svinchuk4
1–4 Russian Technological University MIREA (Moscow, Russia)

1 svyatkin.s.s@edu.mirea.ru, 2 vishnevskij@mirea.ru, 3 novikov_ik@mirea.ru, 4 svinchuk@mirea.ru

The issues of implementation of a no-dependency neural network in CUDA and its practical application in machine learning studies and processing highly sensitive data are discussed. A rise of artificial neural networks has led to a significant increase in demand for qualified ML-engineers and public involvement in discovering inner workings of algorithms that power nearly all AI tech. Despite modern machine learning methods making a huge leap forward compared to as recent as a few years ago, most universities still rely on outdated study materials, which negatively impacts professional skills acquired by students. Moreover, since nowadays neural networks are actively used for processing data, which may contain valuable personal information, state secrets etc., it is imperative that all technologies are certified to guarantee data safety and prevent any leakages.

A fully connected neural network was implemented using CUDA C++, with all steps firmly documented and a step-by-step example of a forward and backward with set numbers is listed. A series of test was conducted to document a final accuracy of up to 99.3% on a train set and 97.93% on a validation set using the MNIST Digits dataset. Performance tests with image size larger than 256x144 have shown a processing speed increase of up to 20 times compared to raw C++ or python implementation.

The results of the study can be applied in developing modern materials for teaching machine learning algorithms to beginning students or used as a base for developing a secure machine learning library for private use in confidential projects.

Svyatkin S.S., Vishnevsky A.K., Novikov I.K., Svinchuk I.V. Feature analysis and implementation of a neural network using NVIDIA CUDA technology. Neurocomputers. 2026. V. 28. № 2. P. 100–110. DOI: https://doi.org/10.18127/j19998554-202602-09 (in Russian)

1. Rashid T. Sozdaem nejronnuyu set': Per. s angl. SPb.: OOO «Al'fa-kniga». 2017. (in Russian).
2. Vakulenko S.A., Zhihareva A.A. Prakticheskij kurs po nejronnym setyam. SPb.: Universitet ITMO. 2018. (in Russian)
3. Grizik F. Prostejshaya nejroset': eshche raz i podrobnee. Habr. URL: https://habr.com/ru/articles/714988/ (data obrashcheniya: 14.01.2025) (in Russian).
4. Ivan'ko A.F., Ivan'ko M.A., Sizova Yu.A. Nejronnye seti: obshchie tekhnologicheskie harakteristiki. Nauchnoe obozrenie. Tekhnicheskie nauki. 2019. № 2. S. 17–23 (in Russian).
5. Tumakov D.N., Giniyatova D.H., Gus'kov V.S. Programmirovanie graficheskih ustrojstv. Tekhnologiya CUDA: Ucheb. posobie. Elektron. tekstovye dan. (1 fajl: 1,70 Mb). Kazan': Izd-vo Kazanskogo un-ta. 2022. (in Russian).
6. Kuzenkov R.S. Analiz effektivnosti raspoznavaniya rukopisnyh simvolov naibolee populyarnymi metodami kontroliruemogo mashinnogo obucheniya. Nauka, tekhnika i obrazovanie. 2017. №6 (36). URL: https://cyberleninka.ru/article/n/analiz-effektivnosti-raspoznavaniya-rukopisnyh-simvolov-naibolee-populyarnymi-metodami-kontroliruemogo-mashinnogo-obucheniya (data obrashcheniya: 10.09.2025) (in Russian).
7. Hizhnyak N. Dolya processorov i videokart AMD rezko podskochila v statistike Steam, a Windows 11 snova stala samoj populyarnoj OS. 3D News: [sajt]. URL: https://3dnews.ru/1120671/dolya-protsessorov-i-videokart-amd-rezko-podskochila-v-statistike-steam-a-windows-11-snova-stala-samoy-populyarnoy-os (data obrashcheniya: 25.08.2025) (in Russian).
8. Volya E.S. Aktual'nye issledovaniya pedagogicheskogo obrazovaniya po rezul'tatam monitoringa aktual'nyh issledovanij v oblasti pedagogicheskogo obrazovaniya: nauchno-issledovatel'skih rabot, razmeshchennyh v baze citirovaniya RINC; dissertacionnyh issledovanij v oblasti pedagogicheskogo obrazovaniya. 2025, sentyabr'. 2025. S. 3–5. (in Russian).
9. Ismagulov M.E. podgotovka dannyh dataseta Crohme 2019 dlya obucheniya nejronnoj seti. Vestnik YuGU. 2024. № 3. URL: https://cyberleninka.ru/article/n/podgotovka-dannyh-dataseta-crohme-2019-dlya-obucheniya-neyronnoy-seti (data obrashcheniya: 12.09.2025) (in Russian).
10. Svidetel'stvo o gosudarstvennoj registracii programmy dlya EVM №2025664828. Nejronnaya set' s apparatnym uskoreniem Nvidia CUDA: programma dlya EVM. A.K. Vishnevskij, A.A. Mryasova, I.V. Svinchuk, S.S. Svyatkin, A.A Solov'eva. 2025 (in Russian).
11. Dobrina M.V. Metody analiza dannyh s ispol'zovaniem iskusstvennyh nejronnyh setej. Nejrokomp'yutery: razrabotka, primenenie. 2023. № 4. S. 45–53 (in Russian).
12. Artem'ev B.V., Vlasov, A.I., Isroilov Zh.O. Analiz programmnyh bibliotek dlya razrabotki vstraivaemyh nejrosetevyh prilozhenij. Nejrokomp'yutery: razrabotka, primenenie. 2023. № 6. S. 5–12 (in Russian).