Adaptive Structure Evolution Convolutional Neural Network for Image Recognition

A new method based on adaptive evolution strategy to automatically search for the optimal feature maps connections between the sub-sampling layer and the convolutional layer of convolutional neural networks without experience is proposed. The purpose of this method is to remove redundant connections in discrete spaces, reduce the complexity of CNN models, and improve the accuracy of image recognition to a certain extent, where the adaptive algorithm can adjust the crossover probability and mutation probability according to the evolution process. The optimal CNN by this method compare with CNN in originally connecting structures and existing methods in the Fashion-MNIST database and MIAS medical image database, this method can effectively optimize the connecting structure, accelerate convergence and improve recognition accuracy.