卷 23, 编号 4 (2024)

The development of digital transformation requires the widespread use of digital technologies in standardization documents. One of the goals is to create standards with machine-understandable content that will allow the use of digital documents at various stages of development and production without the need for a human operator. The purpose of this work is to describe an approach for creating and translating industry normative documents into a machine-understandable representation for their further use in software services and systems. There are three types of SMART standard content: machine-readable, machine-interpretable, and machine-understandable. Knowledge graphs are actively used to formalize data and knowledge when solving various problems. The new two-level approach is proposed for the creation and translation into a machine-understandable representation of regulatory documents as knowledge graphs. The approach defines two types of interpretation of a smart document (human readability and machine understandability) through two related formats: a graph, each semantic node of which represents text in a natural language, and a network of concepts and strict connections. Each node of a human-readable graph corresponds (in general) to a subtree of a machine-readable knowledge graph. As the basis for ensuring the transformation of one form of smart standard representation into another form, LLM models are used, supplemented by a specialized adapter obtained as a result of additional training using the Parameter-Efficient Fine-Tuning approach. Requirements have been established for a set of problem- and subject-oriented tools for generating knowledge graphs. The conceptual architecture of the system for supporting the solution of a set of problems based on knowledge graphs is shown, and the principles for implementing software components that work with smart knowledge for intelligent software services are established.

The problem of training deep neural networks on small samples is especially relevant for medical issues. The paper examines the impact of pixel-wise marking of significant objects in the image, over the true class label, on the quality of the classification. To achieve better classification results on small samples, we propose a multitasking architecture – Unet-boosted classifier (UBC), that is trained simultaneously to solve classification and semantic segmentation problems. As the exploratory dataset, MRI images of patients with benign glioma and glioblastoma taken from the BRaTS 2019 data set are used. One horizontal slice of the MRI image containing a glioma is considered as the input (a total of 380 frames in the training set), and the probability of glioblastoma – as the output. Resnet34 was chosen as the baseline, trained without augmentations with a loss function based on cross-entropy. As an alternative solution, UBC-resnet34 is used – the same resnet34, boosted by a decoder built on the U-Net principle and predicting the pixels with glioma. The smoothed Sorensen-Dice coefficient (DiceLoss) is used as a decoder loss function. Results on the test sample: accuracy for the baseline reached 0.71, for the proposed model – 0.81, and the Dice score – 0.77. Thus, a deep model can be well trained even on a small data set, using the proposed architecture, provided that marking of the affected tissues in the form of a semantic mask is provided.

In recent years, interest in artificial intelligence based on neural network approaches has grown significantly. A number of significant scientific results have been obtained that have found wide application in practice. Generative adversarial neural network models, neural network transformers, and other solutions have attracted much attention. Obvious progress has been achieved in neural network recognition and image generation, text and speech processing, event forecasting, and control of processes that are difficult to formalize. However, it has not yet been possible to endow neural network machines with thinking. All results obtained using neural network machines can be attributed to solutions based on various types of signal binding without full control of their processing processes. Typical representatives of such machines are ChatGPT. The capabilities for intelligently operating various signals in known neural network machines are very limited. Among the main reasons for such limitations, one should highlight the imperfection of the basic principles of neural network information processing used. The properties of neurons have long been considered in a simplified manner. This was due to both gaps in the field of biological research and the lack of opportunities to build large neural networks on complex neuron models. In recent years the situation has changed. New ways to implement large neural networks have emerged. It has also been established that even individual neurons can have extensive internal memory and implement various functions. However, many mechanisms of neuron functioning and their interactions still remain unclear. The issues of controlled associative access to the internal memory of neurons have been little studied. These shortcomings significantly hinder the creation of thinking neural network machines. The object of research in the article is the process of intelligent neural network information processing. The subject of research: principles, models, and methods of such processing. The goal is to expand the functionality of neural network machines to solve difficult-to-formalize creative problems through the development of new principles, models, and methods of intelligent information processing. In the interests of achieving this goal, the operating principles of intelligent neural network machines are clarified, and new models and methods of neural network information processing are proposed. A new model of a pulse neuron is revealed as a basic element of such machines. It is recommended to form the artificial brain of neural network machines in the form of multilayer neural networks endowed with logical structures with neurons of different parameters. A new method of multi-level intelligent information processing in neural network machines based on smart impulse neurons is proposed. The mechanisms of thinking of neural network machines, and the underlying functions of intellectual operation of images and concepts in neural network memory are explained. Simulation results are presented that confirm the validity of the proposed solutions.

The paper describes a general conception of collaborative decision support systems, in which teams providing decision support a) are formed flexibly in accordance with the problem and b) consist of both human experts and intelligent agents implementing AI methods and techniques. An analysis of the key problems of creating collaborative decision support systems based on the collaboration of humans and AI is carried out, the following problems are highlighted: ensuring interoperability (mutual understanding) between heterogeneous team members, reconciling differing positions of participants, ensuring trust between participants, ensuring the effectiveness of joint actions planning and maintaining a balance between predefined workflows and self-organization. Principles for constructing such systems have been formed, offering solutions to the identified problems. In particular, it is proposed to employ an ontology-oriented representation of information about the problem (in the form of multi-aspect ontology), a set of methods for monitoring team activities, reputation scheme, elements of explainable AI, as well as mechanisms of limited self-organization. The proposed concept forms the basis of a software platform for the development of collaborative decision support systems, the main architectural provisions of which are also presented in the paper. The use of the platform is illustrated by an example from the field of rational management of road infrastructure and the creation of a collaborative DSS for the development of measures to reduce road accidents.

A method and a prototype of the program for detecting the presence of birds in the video data flow in real time are presented in the paper. The method is based on the cascade classifier solving the problem of bird detection and identification with the use of a bioacoustic bird scaring system deployed at the Tomsk airport. In our research, the Viola-Jones cascade classifier representing one of the implementations of the Haar cascade algorithm has been used. This algorithm allows objects to be detected in images and videos with high accuracy and rate. In this case, the classifier was leaned on the data set containing images of birds that allowed us to reach high accuracy of bird detection and identification in the videos. The possibilities of the developed classifier are also estimated, and its high productivity is shown. In this study, various methods of machine learning and video data analysis are used to obtain exact and reliable results. As a whole, the present work is an innovative approach to a solution to the urgent problem of airport protection from birds. The application of the developed method has allowed the operating efficiency of the bioacoustic bird scaring system to be increased together with the safety of flights at the Tomsk airport, thereby decreasing the probability of airplane collisions with birds. The novelty of the work consists of the application of the Viola–Jones method for solving the problem of bird detection and identification and estimating its efficiency. Thus, this work is an important contribution to the development of methods for detecting and identifying objects in videos and can also be used in other fields of science and technology in which the automatic detection and classification of objects in the video data flow is required.