To address the problems that the gradient descent method is easy to converge to the local optimum and the convergence speed is slow under large sample data sets, a dynamic attenuation network and a dynamic attenuation gradient descent algorithm are proposed by changing the network structure and gradient descent process in the paper. On the basis of the existing network, an attenuation weight is added between each neuron of each two layers, while an attenuation weight term is introduced in the gradient descent process. The attenuation weight value decreases continuously with iteration, and eventually converges to 0. Due to the addition of the attenuation weight term, the gradient descent speed and convergence speed can be accelerated in the early stage of gradient descent. At the same time, it can avoid crossing over the optimal solution and oscillating around the optimal solution. At the last, it can also improve the probability of the network to obtain the optimal solution. The experimental results on MNIST, CIFAR-10 and CIFAR-100 datasets show that the proposed dynamic attenuation network and dynamic attenuation gradient descent algorithm, compared with the original network that used Adam optimizer and stochastic gradient descent with momentum, improve the test accuracy by 0. 2% ~ 1. 89% and 0. 75% ~ 2. 34%, respectively, while having a faster convergence speed.