https://elib.bsu.by:443/handle/123456789/305993 Proceedings of the 16 th International Conference, Belarus, Minsk, October 17–19, 2023 Tue, 21 Apr 2026 05:56:00 GMT 2026-04-21T05:56:00Z https://elib.bsu.by:443/retrieve/695401/1.jpg https://elib.bsu.by:443/handle/123456789/305993 https://elib.bsu.by:443/handle/123456789/306266 Заглавие документа: Simulation Modelling for Machine Learning Identification of Single Nucleotide Polymorphisms in Human Genomes Авторы: Yatskou, Mikalai M.; Smolyakova, Elizabeth V.; Skakun, Victor V.; Grinev, Vasily V. Аннотация: An approach for simulation modelling of Single Nucleotide Polymorphisms (SNPs) in DNA sequences is proposed, which implements the generation of random events according to the beta or normal distributions, the parameters of which are estimated from the available experimental data. This approach improves the accuracy of determining SNPs in DNA molecules. The verification of the developed model and analysis methods was carried out on a set of reference data provided by the GIAB consortium. The best results were obtained for the machine learning model of Conditional Inference Trees – the accuracy of the SNP identification by the score F1 is 82,8 %, which is higher than those obtained by traditional SNP identification methods, such as binomial distribution, entropy-based and Fisher's exact tests Sun, 01 Jan 2023 00:00:00 GMT https://elib.bsu.by:443/handle/123456789/306266 2023-01-01T00:00:00Z https://elib.bsu.by:443/handle/123456789/306265 Заглавие документа: Neural Networks Interpretation Improvement Авторы: Kroshchanka, Aliaksandr; Golovko, Vladimir Аннотация: The paper is devoted to studying the issues of interpretability of neural network models. Particular attention is paid to the training of heavy models with a large number of parameters. A generalized approach for pretraining deep models is proposed, which allows achieving better performance in final accuracy and interpreting the model output and can be used when training on small datasets. The effectiveness of the proposed approach is demonstrated on examples of training deep neural network models using the MNIST dataset. The obtained results can be used to train fully connected type of layers and other types of layers after applying of flatting operation Sun, 01 Jan 2023 00:00:00 GMT https://elib.bsu.by:443/handle/123456789/306265 2023-01-01T00:00:00Z https://elib.bsu.by:443/handle/123456789/306262 Заглавие документа: Logical-Linguistic and Logical-Probabilistic Methods of Image Classification in Decision-Making Авторы: Gorodetskiy, Andrey E.; Tarasova, Irina L.; Krasavtseva, Anna R. Аннотация: The formation of images in the environment of choice and their classification is one of the important features that characterize the intelligence of modern robots. To do this, we are looking for logical patterns that can explain the available facts and predict the images being formed. Existing neural network methods require the use of pre-training on some training sample. Therefore, objects that are not included in the training sample cannot be classified. Also, the presence of contradictory examples in the training sample and a large noise level in the classified image has a significant impact on the decrease in classification accuracy. Purpose: Construction of new methods for searching for a set of logical connections inherent in the image, construction of classification models and development of structural and linguistic methods of classification of analyzed images. Methods: Logical-linguistic and logical-probabilistic classification methods are proposed, in which the decisive rule of classification is based on calculating the minimum sum of the squares of the differences in the values of the membership functions or probabilities of the elements of the attribute strings of reference and classified images. At the same time, to increase the accuracy of classification, the specified values of membership functions or probabilities can be multiplied by the coefficients of the significance of attributes. Result: The proposed classification algorithms were tested using computer simulation of classification using examples of image recognition in unmanned aerial vehicles, accident risk assessments when driving unmanned vehicles and risk assessments of project financing. The results of computer modeling showed that at a noise level of about 35% - 40%, the accuracy of image classification lay in the range of 78% - 95%. Practical significance: the research results can be used in various intelligent systems to improve the accuracy and speed of image classification Sun, 01 Jan 2023 00:00:00 GMT https://elib.bsu.by:443/handle/123456789/306262 2023-01-01T00:00:00Z https://elib.bsu.by:443/handle/123456789/306259 Заглавие документа: Iterative Selection of Essential Input Features under Conditions of their Multicollinearity in Space Weather Time Series Forecasting Авторы: Shchurov, Nikolay; Isaev, Igor; Barinov, Oleg; Myagkova, Irina; Dolenko, Sergei Аннотация: The paper presents a method for selecting essential input features when predicting the geomagnetic Dst index, based on iterative selection of features with the highest correlation with respect to the target variable and exclusion of features with high cross-correlation. The models were trained on data from October 1997 to 2017. The criterion for the quality of the forecast using selected features was the root-mean squared error of the Dst index forecast based on the selected set of features on independent data (2018-2022) Sun, 01 Jan 2023 00:00:00 GMT https://elib.bsu.by:443/handle/123456789/306259 2023-01-01T00:00:00Z