Автоматический подбор гиперпараметров нейронной сети методом кригинга

Abstract

A choice of optimal hyperparameters requires during a deep neural network training for reaching of the loss function minimum value on a test or validation dataset. However, it is possible to automate this process by algorithms minimizing the value of a loss function. In this paper we perform searching of hyperparameters for deep neural networks using kriging, that is a regression based on Gaussian processes.