Neural network optimizers for buildings architectural style classification task

Abstract

The article considers the problem of classification of images with close interclass relationships - the classification of architectural styles of buildings, where a large number of basic features and characteristics are common to several classes. The object of the research is a neural network learning algorithm for recognizing architectural styles of buildings. A proposed method of neural network training achieves higher accuracy. The proposed optimizer algorithm was implemented and tested to compare the efficiency of the algorithm.
Analysis of existing algorithms and methods has shown shortcomings that are critical to solving the problem of image classification. The advanced algorithm solves some of the shortcomings of existing methods: it does not require a large number of calculations, as the sum of gradients is recursively defined as the average of all past square gradients; the learning speed is adaptive which allows to accelerate learning when the gradient of the function does not change quickly; the algorithm can be used to calculate a group of observations, which will reduce the requirements for memory and the number of calculations; the algorithm is suitable for deep networks; the method regulates the values of the weights of the neural network, not allowing them to increase significantly, which makes the calculations as effective as in SGD for shallow networks; the method takes into account the requirements for the values of the weights of the neural network - they must be small to avoid retraining.
The proposed method of learning the neural network for the classification of archi-?ectural styles of buildings achieves greater accuracy in less learning time. On average, the algorithm showed 1.5 times less value of the loss function at the end of training and 1.13 times faster than using other algorithms.