Перегляд за автором "Kameniuk, N."

Зараз показано 1 - 2 з 2
  • Зображення мініатюри
    Публікація
    A neural network approach for the auto matic selection of a complex of rehabilitation exercises
    (ХНУРЭ, 2021) Butsenko, M. O.; Afanasieva, I. V.; Golian, N. V.; Kameniuk, N.
    This article is devoted to solving the problem of automatic selection of a set of rehabilitation exercises during injuries, considering the state of the human cardiovascular system through the use of neural networks. To solve this problem, it was necessary to choose one of two classical approaches – multiclass classification or multilabel classification, each of which solves the problem of data classification through its own algorithm, and use the selected neural network architecture to create a software system. While working on this system, it was also necessary to solve certain problems related to each of these approaches (the need for a large sample due to the large number of exercises that the system should recommend) or a specific approach (inability to select multiple exercises at once – for Multiclass Classification, lower productivity and the number of supported programming languages – for Multilabel Classification). Samples of different sizes (from 1 million records and more) were used to train the neural network, which were generated through a self-written program that generated a given number of records and wrote them to a .CSV (commaseparated values) file.
  • Зображення мініатюри
    Публікація
    Usage of lstm models for natural language understanding
    (ХНУРЕ, 2021) Daniiel, Y.; Onyshchenko, K.; Kameniuk, N.
    The problem of emotion classification is a complex task of language interpretation. In this work, a number of existing solutions for emotional classification problem were considered. The evaluation of performance of the considered models was conducted. The model for emotion classification in three-sentence conversations is proposed in this work. The model is based on smileys and word embeddings with domain specificity in state of art conversations on the Internet. The model performance is evaluated and compared with language processing model BERT. The proposed model is better at classifying emotions than BERT (F1 78 versus 75). However, modern performance of models for language representation did not achieve the human performance due to the complexity of natural language. There is a variety of factors to consider when choosing the word embeddings and training methods to design the model architecture.