No 3 (2024)
Machine Learning, Neural Networks
Causes of Content Distortion: Analysis and Classification of Hallucinations in Large GPT Language Models
Abstract
The article examines hallucinations produced by two versions of the GPT large language model – GPT-3.5-turbo and GPT-4. The primary aim of the study is to investigate the possible origins and classification of hallucinations as well as to develop strategies to address them. The work reveals the existing challenges that can lead to the generation of content that doesn't correspond to the factual data and misleads users. Detection and elimination of hallucinations play an important role in the development of artificial intelligence by improving natural language processing capabilities. The results of the study have practical relevance for developers and users of language models, due to the provided approaches that improve the quality and reliability of the generated content.
32-41
Multilayer Artificial Neural Networks with S-Parabola Activation Function and their Applications
Abstract
An analysis of modern work in the field of building fast-acting neurons and neural networks was carried out. The algorithm for setting up a multilayer neural network of direct propagation with the activation function of the type "s-parabola" is presented. The setting was carried out based on the method of reverse error propagation, adapted for the specified new function. Examples of using s-parabola in artificial neural networks for solving problems of time series recognition and prediction are considered. Recognition was carried out on the example of typical domestic aircraft, where the objects overall dimensions and the invariant moments of their profiles were used as signs. To predict the time series, the readings of one of the small spacecraft sensors were applied. The solutions quality obtained by the proposed approach was compared with solutions based on neural networks with a traditional "sigmoid". The s-parabola advantage in terms of learning speed and subsequent solution of the applied problem is shown.
42-53
Algorithm for Calculating the Weight Values of a Convolutional Neural Network without Training
Abstract
This study provides a description of the algorithm on the basis of which weights and thresholds are analytically calculated, as well as the number of channels in the layers of a convolutional neural network. Based on the proposed algorithm, a number of experiments were carried out with recognition of the MNIST database. The results of the experiments described in the work showed that the time for calculating the weights of convolutional neural networks is relatively short and amounts to fractions of a second or a minute. The experimental results also showed that already using only 10 selected images from the MNIST database, analytically calculated convolutional neural networks are able to recognize more than half of the images of the MNIST test database, without using neural network training algorithms. Preliminary analytical calculation of the value of the weights of a convolutional neural network allows to speed up the training procedure of a convolutional neural network.
54-70
Analysis of Textual and Graphical Information
Algorithm for Constructing an Oriented Acyclic Graph of Words
Abstract
The algorithm presented in the article makes it possible to efficiently build and modify minimal deterministic finite automata for recognizing a given set of words, including when processing a large amount of information in real time. The key feature of the algorithm is the ability to add new words to the machine and its subsequent minimization on the fly. The algorithm is based on lexicographic ordering of a set of input words and has a low computational complexity compared to traditional algorithms such as the Hopcroft algorithm or an algorithm using the construction of pairs of distinguishable states. The development of this algorithm is aimed at increasing the speed of constructing minimal deterministic finite automata and their modification for effective natural language processing and real-time web content analysis.
104-112
Nomenclature Names Extraction from English and Russian-Language Scientific and Technical Texts
Abstract
The article proposes a method of extracting English and Russian-language nomenclature names from scientific and technical texts on the basis of their structural models. It is noted that nowadays a large number of approaches, methods and software tools for automatic processing of terminological units in natural language texts have been developed, but they do not take into account nomenclature names as a special class of special vocabulary. Their structural and semantic features are analyzed, and on the basis of the analysis models of English and Russian-language nomenclature names are created. A method of automatic extraction of nomenclature names from English and Russian-language scientific and technical texts is proposed. The results of the research can be used in the development of various systems of processing scientific and technical texts, markup of special corpuses, collection of linguistic material in the creation of terminological dictionaries and databases by taking into account a larger number of models of special vocabulary and the use of methods of processing scientific and technical texts in Russian and English.
113-121
Intelligent Planning and Control
Accurate Multiclass Fire Segmentation: Approaches, Neural Networks, and Segmentation Schemes
Abstract
The paper presents a solution to the problem of multiclass flame segmentation with separation by combustion color. The mathematical problems of partial (without separation of the background class into a separate component of the search vector) and full (with separation) segmentation are formulated. A comparison of convolutional neural network methods of UNet, Deeplab and their modern variations, including the wUUNet method developed specifically for the problem under consideration, is carried out. The paper emphasizes the influence of the size of the computation matrix of segmentation computations with the original frame. Both lossy (compressing the frame to the size of the computation matrix and then decompressing it into the original frame) and lossless (applying a single-window frame sizing scheme or multi-window schemes for partitioning the frame into a grid of sub-areas) segmentation schemes are proposed. The best segmentation methods and schemes in terms of quality are selected.
71-86
Predictive Analytics System for the Technical Condition of a Sinter Extractor Using Artificial Intelligence Methods
Abstract
The article describes approaches to building a software system that allows predicting possible failures and malfunctions of industrial equipment based on data on its condition, which will significantly affect the safety of work and the effective functioning of the enterprise. For the task of predicting equipment failures, a model based on "soft voting" between three algorithms with different approaches to classification is proposed: convolutional neural network, logistic regression and the support vector method. A model based on an isolating forest algorithm and an LSTM-based neural network is proposed to predict failures. A web service has been developed that implements the main functions of a predictive analytics system based on artificial intelligence methods: monitoring the technical condition of the excavators in real time, statistical analysis of malfunctions, fault prediction and model training.
87-103
Knowledge Representation
Trust in Artificial Intelligence Technologies
Abstract
Trust in artificial intelligence is a key factor in the widespread introduction of intelligent technologies into the economy and social sphere. The article discusses various aspects of this problem. This is trust in knowledge and data, in artificial intelligence and machine learning models; risks and limits of applicability of the methods and technologies used; explainability of decisions and humanoriented artificial intelligence; primary validation and secondary validation (verification) of created systems; ChatGPT hallucinations and falsifications; ethical, legal and organizational aspects of the use of artificial intelligence.
3-17
Notional Model of Knowledge
Abstract
The article discusses the well-known methods of knowledge representation and processing. A new model, called notional, is proposed, characterized in that the relations (links) between notions are considered as ordinary notions. The notion is considered as a form of thought expressed by a named set of entities. A concrete notion identifies one entity. An abstract notion is formed from other notions by generalization (union) and association (Cartesian product) them. Declarative knowledge is given by enumerable sets of entities, and procedural knowledge is given by solvable ones, where resolving procedures are expressed by formulas of the pure monadic predicate calculus. The description of the applied language of knowledge representation and processing is given. It is proved that queries to the notional model are executed in polynomial time from the logarithm of the average number of entities in notions. The possibility of unsupervised learning notional models is substantiated. It is shown that the notional model makes it possible to visually represent and effectively process both declarative and procedural knowledge.
18-31