Image recognition method based on deep neural network model parameter modulation

Abstract

The invention relates to an image recognition method based on deep neural network model parameter modulation, and the method comprises the steps of obtaining loss function loss values corresponding toa batch of training samples through forward propagation based on a randomly initialized deep neural network model; utilizing a gradient calculation function to obtain a first order gradient and a second order gradient of the network; calculating the curvature corresponding to each layer of network parameters according to the first-order gradient and the second-order gradient of the network and the loss value of the loss function; carrying out segmented modulation on the curvature corresponding to the network parameters of each layer, then calculating a corresponding curvature radius, and updating the network parameters of each layer; inputting a next batch of training samples, and repeatedly carrying out iterative optimization on the network parameters until the deep neural network modelconverges; and inputting a to-be-identified image into the optimized and trained deep neural network model to obtain an identified image. According to the invention, the problem that the training performance of the existing deep neural network gradually becomes poor along with the increase of the batch number is solved.

Description

technical field [0001] The invention relates to the technical field of deep learning, in particular to an image recognition method based on deep neural network model parameter modulation. Background technique [0002] In the field of deep learning technology, how to effectively update network parameters is one of the core issues. The existing gradient update method is based on the stochastic gradient descent method (SGD), supplemented by a variety of optimization methods, such as Adagrad, RMSprop, Adam, etc. , these methods only use the first-order gradient information of the neural network, and do not consider the second-order gradient information, so the local curvature characteristics of the loss function curve where each network parameter is located cannot be obtained, and all network parameters can only use one same learning rate, the way to update parameters is relatively extensive. The advantage of this network parameter update method is that it is random. When there...

AI Technical Summary

Problems solved by technology

[0003] In view of the above analysis, the present invention aims to provide an image recognition method based on deep neural network model parameter modulation to solve the problem that the training performance of the deep neural network model gradually decreases with the increase of the number of batches under the existing network parameter modulation. Variation problem

Images