An Evolutionary Algorithm-Based Neural Architecture Search for Feature Separation in Classification Problems

In this paper, we propose Evolutionary Algorithm-based Neural Architecture Search(EvoNAS), which use Evolutionary Algorithm to find the optimized Convolutional Neural Network(CNN) and combine with Distribution Estimation to improve efficiency during the search process to improve the classification performance. The optimized CNN reveal the process of classification. It can classify the features according to specified classes, improving the classification accuracy of similar features while achieving the desired classification accuracy. We conduct feature separation experiments On the publicly available CIFAR dataset, we use EvoNAS to conducte feature separation experimentson. EvoNAS is subjected to feature separation experiments on the publicly available CIFAR dataset. The classification accuracy has been improved by 0.55% and 0.56% respectively compare to the original network. In addition, the feature visualization and attribution visualization improve the classificatory interpretability of CNN.