<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE article PUBLIC "-//NLM//DTD JATS (Z39.96) Journal Publishing DTD v1.3 20210610//EN" "JATS-journalpublishing1-3.dtd">
<article article-type="research-article" dtd-version="1.3" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xml:lang="ru"><front><journal-meta><journal-id journal-id-type="publisher-id">ntv</journal-id><journal-title-group><journal-title xml:lang="ru">Научно-технический вестник информационных технологий, механики и оптики</journal-title><trans-title-group xml:lang="en"><trans-title>Scientific and Technical Journal of Information Technologies, Mechanics and Optics</trans-title></trans-title-group></journal-title-group><issn pub-type="ppub">2226-1494</issn><issn pub-type="epub">2500-0373</issn><publisher><publisher-name>Университет ИТМО</publisher-name></publisher></journal-meta><article-meta><article-id pub-id-type="doi">10.17586/2226-1494-2026-26-3-544-555</article-id><article-id custom-type="elpub" pub-id-type="custom">ntv-622</article-id><article-categories><subj-group subj-group-type="heading"><subject>Research Article</subject></subj-group><subj-group subj-group-type="section-heading" xml:lang="ru"><subject>КОМПЬЮТЕРНЫЕ СИСТЕМЫ И ИНФОРМАЦИОННЫЕ ТЕХНОЛОГИИ</subject></subj-group><subj-group subj-group-type="section-heading" xml:lang="en"><subject>COMPUTER SCIENCE</subject></subj-group></article-categories><title-group><article-title>Принцип максимальной энтропии в глубоком обучении: от теории к практическому регуляризатору для сложных данных</article-title><trans-title-group xml:lang="en"><trans-title>Maximum entropy principle in deep learning: from theory to a practical regularizer for complex data</trans-title></trans-title-group></title-group><contrib-group><contrib contrib-type="author" corresp="yes"><contrib-id contrib-id-type="orcid">https://orcid.org/0000-0001-7212-5230</contrib-id><name-alternatives><name name-style="eastern" xml:lang="ru"><surname>Тимофеев</surname><given-names>А. В.</given-names></name><name name-style="western" xml:lang="en"><surname>Timofeev</surname><given-names>A. V.</given-names></name></name-alternatives><bio xml:lang="ru"><p>Тимофеев Андрей Владимирович — доктор технических наук, научный директор</p><p>sc 56689367600</p><p>Санкт-Петербург, 197341</p></bio><bio xml:lang="en"><p>Andrey V. Timofeev — D.Sc., Chief Scientific Officer</p><p>Saint Petersburg, 197341</p></bio><email xlink:type="simple">timofeev.andrey@gmail.com</email><xref ref-type="aff" rid="aff-1"/></contrib></contrib-group><aff-alternatives id="aff-1"><aff xml:lang="ru"><institution>ООО «Центр ИИ для Стран ШОС+»</institution><country>Россия</country></aff><aff xml:lang="en"><institution>LLC “AI Center for SCO+ Countries”</institution><country>Russian Federation</country></aff></aff-alternatives><pub-date pub-type="collection"><year>2026</year></pub-date><pub-date pub-type="epub"><day>09</day><month>07</month><year>2026</year></pub-date><volume>26</volume><issue>3</issue><fpage>544</fpage><lpage>555</lpage><permissions><copyright-statement>Copyright &amp;#x00A9; Тимофеев А.В., 2026</copyright-statement><copyright-year>2026</copyright-year><copyright-holder xml:lang="ru">Тимофеев А.В.</copyright-holder><copyright-holder xml:lang="en">Timofeev A.V.</copyright-holder><license xml:lang="ru" license-type="creative-commons-attribution" xlink:href="https://creativecommons.org/licenses/by/4.0/" xlink:type="simple"><license-p>Данная работа распространяется под лицензией Creative Commons Attribution 4.0.</license-p></license><license xml:lang="en" license-type="creative-commons-attribution" xlink:href="https://creativecommons.org/licenses/by/4.0/" xlink:type="simple"><license-p>This work is licensed under a Creative Commons Attribution 4.0 License.</license-p></license></permissions><self-uri xlink:href="https://ntv.elpub.ru/jour/article/view/622">https://ntv.elpub.ru/jour/article/view/622</self-uri><abstract><p>Введение. Представлена и экспериментально проверена гипотеза о том, что в машинном обучении максимизация энтропии распределения эмбеддингов при заданных ограничениях ведет к улучшению обобщающей способности моделей. Метод. Для реализации гипотезы предложен новый метод — регуляризация методом максимизации энтропии на гиперсфере с разделением классов, предназначенный для задач классификации с высокой зашумленностью и дисбалансом классов (например, распознавание сигналов Orthogonal Frequency Division Multiplexing (OFDM)). Метод основан на расширении функции потерь классической модели за счет трех специализированных компонент: первый максимизирует энтропию через равномерность распределения на гиперсфере, второй контролирует ковариацию признаков и третий обеспечивает межклассовую разделимость. Динамическая регуляризация позволяет адаптивно балансировать влияние энтропийного члена в процессе обучения. Основные результаты. Показана связь между максимизацией энтропии распределения эмбеддингов и сужением границ обобщения модели. Экспериментально на задаче классификации OFDM-сигналов продемонстрировано, что предложенный метод обеспечивает формирование высокоравномерного пространства признаков: среднее косинусное сходство снижается до 0,064, покрытие достигает 1,71, энтропия распределения сходств — 10,34 нат при сохранении отличной дискриминативности, что подтверждается точностью 0,97 и сбалансированным Macro F1-score 0,98. Метод эффективно устраняет переобучение и повышает устойчивость модели на миноритарных классах. Обсуждение. Разработанный подход, явно максимизирующий энтропию представлений при заданных ограничениях, предоставляет перспективный принцип для улучшения обобщающей способности в условиях недостаточности или зашумленности данных. Перспективы применения метода включают ослабление теоретических предположений и применение в смежных областях, где надежность классификации критически зависит от качества представлений.</p></abstract><trans-abstract xml:lang="en"><p>We hypothesize and experimentally validate that maximizing the entropy of embedding distributions under specific constraints enhances the generalization ability of machine learning models. To test this hypothesis, we propose a novel regularization method — Hyperspherical Entropy Maximization with Class Separation — designed for noisy and imbalanced classification tasks (e.g., Orthogonal Frequency Division Multiplexing (OFDM) signal recognition).The method is based on extending the loss function of the classical model with three specialized components: the first maximizes entropy by enforcing uniformity of distribution on the hypersphere, the second controls the covariance of features, and the third ensures inter-class separability. Dynamic regularization allows adaptively balancing the influence of the entropy term during training. We establish a link between maximizing the entropy of the embedding distribution and a reduction in the model generalization error bound. In experiments on an OFDM signal classification task, our method shapes a highly uniform feature space: mean cosine similarity decreases to 0.064, coverage reaches 1.71, and the entropy of the similarity distribution attains 10.34 nats, while preserving strong discriminative power, achieving 0.97 accuracy and a Macro F1-score of 0.98, outperforming standard regularization techniques. The approach effectively mitigates overfitting and improves robustness on minority classes. Our approach, which explicitly maximizes representation entropy under task-specific constraints, establishes a promising principle for enhancing generalization in conditions of data scarcity or high noise. Future work will focus on relaxing theoretical assumptions and applying the method to other domains where classification reliability critically depends on representation quality.</p></trans-abstract><kwd-group xml:lang="ru"><kwd>эмбеддинг</kwd><kwd>машинное обучение</kwd><kwd>обобщающая способность</kwd><kwd>OFDM</kwd><kwd>энтропия</kwd><kwd>регуляризация</kwd></kwd-group><kwd-group xml:lang="en"><kwd>embeddings</kwd><kwd>machine learning</kwd><kwd>generalization</kwd><kwd>OFDM</kwd><kwd>entropy</kwd><kwd>regularization</kwd></kwd-group></article-meta></front><back><ref-list><title>References</title><ref id="cit1"><label>1</label><citation-alternatives><mixed-citation xml:lang="ru">Bengio Y., Courville A., Vincent P. Representation learning: A review and new perspectives // IEEE Transactions on Pattern Analysis and Machine Intelligence. 2013. V. 35. N 8. P. 1798–1828. doi: 10.1109/tpami.2013.50</mixed-citation><mixed-citation xml:lang="en">Bengio Y., Courville A., Vincent P. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2013, vol. 35, no. 8, pp. 1798–1828. doi: 10.1109/tpami.2013.50</mixed-citation></citation-alternatives></ref><ref id="cit2"><label>2</label><citation-alternatives><mixed-citation xml:lang="ru">Chen T., Kornblith S., Norouzi M., Hinton G. A simple framework for contrastive learning of visual representations // Proc. of the 37th International Conference on Machine Learning. 2020. P. 1597–1607.</mixed-citation><mixed-citation xml:lang="en">Chen T., Kornblith S., Norouzi M., Hinton G. A simple framework for contrastive learning of visual representations. Proc. of the 37th International Conference on Machine Learning, 2020, pp. 1597–1607.</mixed-citation></citation-alternatives></ref><ref id="cit3"><label>3</label><citation-alternatives><mixed-citation xml:lang="ru">He K., Fan H., Wu Y., Xie S., Girshick R. Momentum contrast for unsupervised visual representation learning // Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2020. P. 9726–9735. doi: 10.1109/cvpr42600.2020.00975</mixed-citation><mixed-citation xml:lang="en">He K., Fan H., Wu Y., Xie S., Girshick R. Momentum contrast for unsupervised visual representation learning. Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 9726–9735. doi: 10.1109/cvpr42600.2020.00975</mixed-citation></citation-alternatives></ref><ref id="cit4"><label>4</label><citation-alternatives><mixed-citation xml:lang="ru">Wang T., Isola P. Understanding contrastive representation learning through alignment and uniformity on the hypersphere // Proc. of the 37th International Conference on Machine Learning. 2020. P. 9929-9939.</mixed-citation><mixed-citation xml:lang="en">Wang T., Isola P. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. Proc. of the 37th International Conference on Machine Learning, 2020, pp. 9929-9939.</mixed-citation></citation-alternatives></ref><ref id="cit5"><label>5</label><citation-alternatives><mixed-citation xml:lang="ru">Chakraborty D., LeCun Y., Rudner T.G.J., Learned-Miller E. Improving pre-trained self-supervised embeddings through effective entropy maximization // arXiv. 2024. arXiv:2411.15931. doi: 10.48550/arXiv.2411.15931</mixed-citation><mixed-citation xml:lang="en">Chakraborty D., LeCun Y., Rudner T.G.J., Learned-Miller E. Improving pre-trained self-supervised embeddings through effective entropy maximization. arXiv, 2024. arXiv:2411.15931. doi: 10.48550/arXiv.2411.15931</mixed-citation></citation-alternatives></ref><ref id="cit6"><label>6</label><citation-alternatives><mixed-citation xml:lang="ru">Берестова С.А., Копытов Н.П., Митюшов Е.А. Дискретные ориентации космического аппарата // Инженерный журнал: наука и инновации. 2017. № 7 (67). С. 7. doi: 10.18698/2308-6033-2017-7-1661</mixed-citation><mixed-citation xml:lang="en">Berestova S.A., Kopytov N.P., Mityushov E.A. Spacecraft discrete orientations. Engineering Journal: Science and Innovation, 2017, no. 7 (67), pp. 7. (in Russian). doi: 10.18698/2308-6033-2017-7-1661</mixed-citation></citation-alternatives></ref><ref id="cit7"><label>7</label><citation-alternatives><mixed-citation xml:lang="ru">Pang T., Yang X., Dong Y., Xu K., Zhu J., Su H. Boosting adversarial training with hypersphere embedding // Proc. of the 34th Conference on Neural Information Processing Systems (NeurIPS). 2020.</mixed-citation><mixed-citation xml:lang="en">Pang T., Yang X., Dong Y., Xu K., Zhu J., Su H. Boosting adversarial training with hypersphere embedding. Proc. of the 34th Conference on Neural Information Processing Systems (NeurIPS), 2020.</mixed-citation></citation-alternatives></ref><ref id="cit8"><label>8</label><citation-alternatives><mixed-citation xml:lang="ru">Jakovác A., Telcs A. A note on representational understanding // Entropy. 2022. V. 24. N 9. P. 1313. doi: 10.3390/e24091313</mixed-citation><mixed-citation xml:lang="en">Jakovác A., Telcs A. A note on representational understanding. Entropy, 2022, vol. 24, no. 9, pp. 1313. doi: 10.3390/e24091313</mixed-citation></citation-alternatives></ref><ref id="cit9"><label>9</label><citation-alternatives><mixed-citation xml:lang="ru">Garg V.K., Jegelka S., Jaakkola T. Generalization and representational limits of graph neural networks // Proc. of the 37th International Conference on Machine Learning. 2020. P. 3419–3430.</mixed-citation><mixed-citation xml:lang="en">Garg V.K., Jegelka S., Jaakkola T. Generalization and representational limits of graph neural networks. Proc. of the 37th International Conference on Machine Learning, 2020, pp. 3419–3430.</mixed-citation></citation-alternatives></ref><ref id="cit10"><label>10</label><citation-alternatives><mixed-citation xml:lang="ru">Ren B., Teh K.C., An H., Gunawan E. OFDM modulation classification using Cross-SKNet with blind IQ imbalance and carrier frequency offset compensation // IEEE Transactions on Vehicular Technology. 2024. V. 73. N 6. P. 8389–8403. doi: 10.1109/tvt.2024.3356606</mixed-citation><mixed-citation xml:lang="en">Ren B., Teh K.C., An H., Gunawan E. OFDM modulation classification using Cross-SKNet with blind IQ imbalance and carrier frequency offset compensation. IEEE Transactions on Vehicular Technology, 2024, vol. 73, no. 6, pp. 8389–8403. doi: 10.1109/tvt.2024.3356606</mixed-citation></citation-alternatives></ref><ref id="cit11"><label>11</label><citation-alternatives><mixed-citation xml:lang="ru">Данфорд Н., Шварц Дж. Линейные операторы. Т. 1. М.: Иностранная литература, 1966. С. 326–330.</mixed-citation><mixed-citation xml:lang="en">Dunford N., Schwartz J. Linear Operators. Part 1. Interscience Publishers, 1958, pp. 326–330</mixed-citation></citation-alternatives></ref><ref id="cit12"><label>12</label><citation-alternatives><mixed-citation xml:lang="ru">Рудин У. Основы математического анализа. М.: Мир, 1976. С. 99.</mixed-citation><mixed-citation xml:lang="en">Rudin W. Principles of Mathematical Analysis. Moscow, Mir Publ., 1976, pp. 99. (in Russian)</mixed-citation></citation-alternatives></ref><ref id="cit13"><label>13</label><citation-alternatives><mixed-citation xml:lang="ru">Anderson Т.W. The integral of a symmetric unimodal function over a symmetric convex set and some probability inequalities // Proceedings of the American Mathematical Society. 1955. V. 6. N 2. P. 170–176. doi: 10.1090/S0002-9939-1955-0069229-1</mixed-citation><mixed-citation xml:lang="en">Anderson Т.W. The integral of a symmetric unimodal function over a symmetric convex set and some probability inequalities. Proceedings of the American Mathematical Society, 1955, vol. 6, no. 2, pp. 170–176. doi: 10.1090/S0002-9939-1955-0069229-1</mixed-citation></citation-alternatives></ref><ref id="cit14"><label>14</label><citation-alternatives><mixed-citation xml:lang="ru">Prekopa A. On logarithmic concave measures and functions // Acta Scientiarum Mathematicarum. 1973. V. 34. P. 335–343.</mixed-citation><mixed-citation xml:lang="en">Prekopa A. On logarithmic concave measures and functions. Acta Scientiarum Mathematicarum, 1973, vol. 34, pp. 335–343.</mixed-citation></citation-alternatives></ref><ref id="cit15"><label>15</label><citation-alternatives><mixed-citation xml:lang="ru">Mohri M., Rostamizadeh A., Talwalkar A. Foundations of Machine Learning. MIT Press, 2018. 504 p.</mixed-citation><mixed-citation xml:lang="en">Mohri M., Rostamizadeh A., Talwalkar A. Foundations of Machine Learning. MIT Press, 2018, 504 p.</mixed-citation></citation-alternatives></ref><ref id="cit16"><label>16</label><citation-alternatives><mixed-citation xml:lang="ru">Bartlett P.L., Mendelson S. Rademacher and Gaussian complexities: Risk bounds and structural results // Journal of Machine Learning Research. 2002. V. 3. P. 463–482.</mixed-citation><mixed-citation xml:lang="en">Bartlett P.L., Mendelson S. Rademacher and Gaussian complexities: Risk bounds and structural results. Journal of Machine Learning Research, 2002, vol. 3, pp. 463–482.</mixed-citation></citation-alternatives></ref><ref id="cit17"><label>17</label><citation-alternatives><mixed-citation xml:lang="ru">Cover T.M., Thomas J.A. Elements of Information Theory. John Wiley &amp; Sons, 2006. 784 p.</mixed-citation><mixed-citation xml:lang="en">Cover T.M., Thomas J.A. Elements of Information Theory. John Wiley &amp; Sons, 2006, 784 p.</mixed-citation></citation-alternatives></ref><ref id="cit18"><label>18</label><citation-alternatives><mixed-citation xml:lang="ru">El-Khamy S.E., Elsayed H.A., Rizk M.M. C45. Classification of OFDM signals using higher order statistics and clustering techniques // Proc. of the 29th National Radio Science Conference (NRSC). 2012. P. 541–549. doi: 10.1109/nrsc.2012.6208563</mixed-citation><mixed-citation xml:lang="en">El-Khamy S.E., Elsayed H.A., Rizk M.M. C45. Classification of OFDM signals using higher order statistics and clustering techniques. Proc. of the 29th National Radio Science Conference (NRSC), 2012, pp. 541–549. doi: 10.1109/nrsc.2012.6208563</mixed-citation></citation-alternatives></ref></ref-list><fn-group><fn fn-type="conflict"><p>The authors declare that there are no conflicts of interest present.</p></fn></fn-group></back></article>
