Application of Modular Neural Networks for Image Recognition in Foggy Computing Environments

Abstract

The paper considers various approaches to the decomposition of artificial neural networks for the purpose of their application on fog computing nodes. Based on the requirements for the organization of fog computing, a method of dividing the input information into subspaces by means of wavelet transform and subsequent proportional division of all layers of the neural network is proposed. The proposed approach achieves a significant gain in the amount of information transferred between modules compared to the currently used layer-by-layer partitioning. In addition, the proposed method optimizes the load on fog computing nodes by partially utilizing the modules.