Багатошарова матрична нейронна мережа у якості альтернативи багатошаровому перцептрону для задач класифікації зображень

Abstract

This paper is devoted to the exploration of the advantages of using matrix neuron based layers as an alternative to fully connected layers in the head of image classification networks. The recent advances in the application of neural networks for computer vision tasks, including visual transformers and full MLP architectures for image processing, are considered. The advantages of using matrix layers, including decreased number of parameters and reducing information loss, are described. It has been found that replacing fully connected layers with matrix layers and a vectorization-free approach has the potential for improving neural networks performance.