Анализ функций потерь при обучении сверточных нейронных сетей с оптимизатором Adam для классификации изображений

Abstract

The problem of analyzing and determining the loss functions used in training convolutional neural networks (CNN), which provide the highest accuracy of image classification, is considered. The loss functions implemented on the basis of the Keras and TensorFlow machine learning libraries, using which CNNs can be designed, trained, and applied in the Python language, are analyzed. The MNIST, EMNIST and CIFAR10 datasets are used as image sources for CNN training and validation purposes. These data sets are used to train CNNs having different architectures, for each of which the history of its training is preserved (information about the accuracy of image classification on training and validation samples after each training epoch). The evaluation results are then generalized. The highest achieved accuracy of image classification makes it possible to justify the expediency of using the relevant loss functions in training the CNNs.