Improved evolutionary neural network architecture search method based on super network

Abstract

The invention relates to an improved evolutionary neural network architecture search method based on a super network. The method comprises the following steps of S1, taking an input layer as a first layer, and packaging five calculation modules; S2, binarizing the connection of the computing nodes in the neural network; S3, learning a structure weight for each computing node, and S4, constructinga parent population P by adopting a binary fancy competition selection method. S5, forming a filial generation population Q; and S6, performing mutation operation on individuals in the filial generation population Q. S7, decoding each individual in the filial generation population Q into a corresponding neural network to obtain a structural weight; and S8, combining the parent population P and thechild population Q into a population R, selecting a plurality of individuals as an original population of the next generation by adopting an environment selection method, and feeding back to the stepS4 until a predetermined maximum evolution algebra is reached, and after the evolution is finished, outputting the individual with the highest fitness value as the optimal neural network architecture.

Description

technical field [0001] The invention relates to the technical field of image classification model construction, in particular to an improved supernetwork-based evolutionary neural network architecture search method. Background technique [0002] The image classification task is an image processing technology that distinguishes different types of targets based on the different feature information reflected in the picture. Since many models applied to image classification tasks can be transferred to other computer vision fields as feature extraction networks, image classification tasks are a basic task in the field of computer vision, and the design of image classification models is also a hot spot for researchers. However, manual design of neural network models requires experienced experts to design neural network models with excellent performance through careful research on the distribution and characteristics of data sets and trial and error. Therefore, huge time and labor...

AI Technical Summary

Problems solved by technology

[0006] Aiming at the insufficiency of the hypernetwork-based neural network architecture search method in the prior art, and the supernetwork training convergence speed is slow or even unable to converge, the technical problem to be solved by the present invention is to provide a supernetwork-based evolutionary neural network architecture A search method that uses evolutionary algorithms as a search strategy to automatically generate neural network architectures based on hypernetworks to improve classification accuracy for image classification tasks

Images