Analysis of modern SOTA-architectures of artificial neural networks for solving problems of image classification and object detection

Abstract

The scientific research is focused on conducting a study of current artificial neural network architectures in order to highlight the advantages and disadvantages of current approaches. The relevance of the research relies on the growing interest in machine learning technologies and regular improvement of computer vision algorithms.Within the scope of this paper, an analytical study of the advantages and disadvantages of existing solutions has been conducted and advanced SOTA architectures have been reviewed. The most effective approaches to improve the accuracy of basic models have been studied. The number of parameters used, the size of the training sample, the accuracy of the model, its size, adaptability, complexity and the required computational resources for training a single architecture were determined.Prospects for further research in the field of hybridization of convolutional neural networks and visual transformers are revealed, and a new solution for building a complex neural network architecture is proposed.In the framework of the present research work, a detailed analysis of the internal structure of the most effective neural network architectures.Plots of the accuracy dependence on the number of parameters used in the model and the size of the training sample are plotted. The conducted comparative analysis of the efficiency of the considered solutions allowed to single out the most effective methods and technologies for designing artificial neural network architectures. A novel method focused on creating a complex adaptive model architecture that can be dynamically tuned depending on an input set of parameters is proposed, representing a potentially significant contribution to the field of adaptive neural network design.