Improving The Performance of Convolutional Neural Networks using Evolutionary Computing

تحسين أداء تجميع الشبكات الطبيعية SNFUL باستخدام الحوسبة المسطحة

Abstract

Convolutional Neural Networks (CNNs) have achieved remarkable success with numerous real-world issues in recent years. The structure of these networks is heavily influenced by a number of parameters, such as the number and type of layers, the size and number of cores, and the type of activation function. In this article, genetic algorithms were used to design CNNs structures, because genetic algorithms are able to apply learning in an automatic way. The algorithm was tested on Cifar10 and Cifar100 datasets, compared with three newly designed CNNs, two competitors designed semi-automatic CNN structures, and others designed fully automatic CNN structures. As the results CIFAR10 Classification error (4.3), Parameters number (M = 106) (2.1 M) and Execution-time (day) (40). CIFAR100 Classification error (20.85), Parameters number (M = 106) (5.5 M) and Execution-time (day) (84). The results also show that the parameters' number of the best structure reached using the proposed algorithm is less than in the automatic algorithms that were compared with it (Block-QNN-S and Large-Scale Evolution). This is within an implementation time of the proposed algorithm of 84 days, according to the computational resources.