Search academic texts

Information × Registration Number 0226U001771, (0124U001734) , R & D reports Title Development of non-connection models of artificial neural networks, methods for their construction and training popup.stage_title Розробка біологічно обґрунтованої моделі нейрона-детектора та архітектури структурно-детекторної нейронної мережі й алгоритму їх навчання Head Dorofieiev Iurii I., д.т.н. Registration Date 30-01-2026 Organization National Technical University "Kharkiv Polytechnic Institute" popup.description1 Increasing the efficiency of solving problems of image classification and clustering based on the application of models, methods and algorithms for building and training structural-detector artificial neural networks popup.description2 The modern development of artificial intelligence (AI) faces three critical barriers: the exponential growth of energy consumption for model training, the need for massive labeled datasets, and the problem of lack of interpretability of decisions. The relevance of this research lies in creating a shift in the current AI paradigm, bringing the operating principles of artificial neural networks (ANNs) closer to biological and physical processes. This enables the development of autonomous systems capable of learning from small datasets, which is critical for defense, medicine, and robotics under conditions of limited resources. The study addresses the fundamental problem of creating self-organizing intelligent systems that do not require statistical learning. This allows bridging the gap between the abstract concept of information and its physical carrier, enabling system learning based on the principle of free energy minimization, similar to processes observed in living nature. The objective of the research is the development and experimental validation of a theoretical framework that postulates the energetic nature of information and defines its role in the processes of self-organization and evolution of complex information systems. The following tasks were accomplished in the work: - A new architecture of the Compartmental Attractor Neural Network. - A learning method without backpropagation was implemented. - Few-Shot Learning was achieved: the model’s ability to learn from extremely small samples (5–6 examples per class) was experimentally confirmed. On the MNIST dataset, an accuracy of ~82% was reached using minimal resources, which is unattainable for classical ANNs under such conditions. - The problem of AI transparency was addressed. Classification is performed through the calculation of the Graph Edit Distance metric, which allows clearly explaining why an object is assigned to a particular class based on the similarity of its geometric structure. Product Description popup.authors Yurii V. Parzhyn Kostiantyn O. Bokhan Mykyta O. Lapin popup.nrat_date 2026-01-30 Close