Examinando por Materia "Convolutional Neural Networks"

Mostrando 1 - 1 de 1
  • No hay miniatura disponible
    Ítem
    Variability modeling language and tool to represent, configure and evaluate Convolutional Neural Network architectures
    (Universidad EAFIT, 2024) Murillo Portocarrero, Julián Alexander; Mazo Peña, Raúl
    The process of designing Convolutional Neural Network (CNN) architectures currently relies on manual design or the involvement of experts. This process is not only time-consuming but also expensive due to the sheer number of combinations that architects need to do to arrive at the right network hyper-parameters that fit the current problem. In this work, we analyze the current state-of-the-art, in which several approaches have been proposed to automate such processes and explore alternatives for the automated design of Convolutional Neural Networks and Neural Architecture Search (NAS). Additionally, this work proposes a method for hyper-parameter variability generation from a variability model of such convolutional neural networks. The variability model proposed in this master thesis is used to represent, intensively, the valid combinations of parameters corresponding to each convolutional neural network. The language, called CNN variability language, borrows some concepts from Software Product Lines (SPL) and was created on the VariaMos platform to enable architects and engineers not just to create CNN architectures but also to automatically generate configurations, generate executable Jupyter Notebooks for each configuration, and generate comparison reports to speed up the NAS process.