Deep neural network optimizing method based on coevolution and back propagation

Abstract

The invention discloses a deep neural network optimizing method based on coevolution and back propagation, belongs to the technical field of combination of deep learning and evolutionary computation, and mainly aims to solve the problem of tendency to fall into a local optimal solution during training of the deep neural network. The deep neural network optimizing method comprises the following steps: (1) firstly, optimizing a network by using a back propagation algorithm; (2) when a stopping condition is satisfied, optimizing by using a co-evolution algorithm; (3) continually repeating the previous steps until an iterative stopping condition is satisfied; and (4) performing final optimization to obtain a weight and a deviation as optical parameters. In the deep neural network optimizing method, the advantages of an evolution algorithm are applied to the training of the deep neural network, and large-scale parameters are optimized by coevolution. Meanwhile, a selection strategy is designed in conjunction with the back propagation algorithm in order to increase the optimizing speed of the coevolution, so that the whole network can be trained more efficiently, high optimizing performance is achieved, and the classification accuracy of the network is increased.

Description

technical field [0001] The invention belongs to the combination of deep learning and evolutionary algorithm field, mainly solves the parameter optimization problem of deep neural network, specifically provides a deep neural network optimization method based on collaborative evolution and back propagation, and realizes the optimization of deep neural network parameters. Background technique [0002] Since the 1980s, the neural network (Neural Network, NN) has entered the fast lane of development. The major breakthrough of new scientific theories and the rapid development of high-performance computers have brought NN back to life. In 1982, Professor Hopfield of California Institute of Technology proposed the famous Hopfield neural network model, which strongly promoted the research of NN. In the mid-1980s, Ackley, Hinton and Sejnowski introduced a random mechanism based on the idea of ​​simulated annealing in the Hopfield NN model, and based on this model analyzed the differen...

AI Technical Summary

Problems solved by technology

[0004] One of the current optimization algorithms for deep neural networks is: gradient-based backpropagation algorithm, which tends to fall into a local optimal solution as the complexity of network results and data increases; the other is evolutionary algorithm, due to With the continuous increase of the parameter dimension, the traditional evolutionary algorithm is no longer applicable

Images