Fuzzy granular deep convolutional network with residual structures

Abstract

In recent years, the deep neural network technology has developed rapidly and has been effective in processing and analyzing images, videos, sounds and many other aspects. However, traditional neural networks only learn the samples themselves and ignore the differences between samples. When the network depth is too deep, network degradation phenomenon may occur. To solve these drawbacks, we propose the fuzzy granular deep convolutional network with residual structures. Firstly, we define the concept, operation and correlation measures of fuzzy granules, and construct the granulation reference system by random sampling for fuzzy granulation. Furthermore, modules of granular neuron, granular activation function, granular convolution, and granular residual are defined, and the fuzzy granular deep convolutional network with residual structures is built. The loss functions and learning algorithm are designed for the granular neural network, and the fuzzy granular neural network is successfully trained. The fuzzy granular neural network owns the characteristics of multi-granularity, multi-angle and structured, which has good generalization performance. Finally, experiments are conducted on the Cifar100 and Tiny Imagenet datasets. The experimental results show the effectiveness of the residual granular neural network. We further found that the granular neural network could alleviate the degradation problem. Moreover granular neural network with fewer hidden layers can achieve the function of a neural network with more hidden layers.