一种基于混合粒子群优化算法的深度卷积神经网络架构搜索方法

神经架构搜索（neural architecture search，NAS）技术自动寻找神经网络中各层的最佳组合和连接方式，以及各种超参数的最佳分布。该方法从搜索空间生成若干不同的卷积神经网络（CNN），使用混合粒子群优化（hybrid particle swarm optimization，HPSO）算法，将一定数目的神经网络个体视做一个群体，将每个网络个体在评价指标下的表现值视做适应度，在给定的世代数范围内，每个神经网络个体都学习自身的历史最佳适应度个体，和整个群体的最佳适应度个体，迭代改善自身的网络架构。实验结果表明，算法运行中出现的最优网络架构，在图像分类任务的多个基准数据集上，与手工设计的神经网络和以遗传算法为基础的NAS算法相比，在网络参数数量和准确率的平衡上取得了有竞争力的结果。

Deep convolutional neural architecture search method based on hybrid particle swarm optimization algorithm

The neural architecture search(NAS) technique automatically finds the optimal combination and connectivity of layers in a neural network, as well as the optimal distribution of various hyperparameters. The method generated a number of different convolutional neural network(CNN) from the search space, and used a hybrid particle swarm optimization(HPSO) algorithm to treat a certain number of neural network individuals as a population, and the performance of each individual under the evaluation metric as the fitness, within a given number of generations. Each neural network individual learnt its own historical best fitness individual, and the best fitness individual of the whole population, and iteratively improved its own network architecture. Experimental results show that the optimal network architecture emerging from the algorithm runs achieves competitive results in terms of the trade-off between the number of network parameters and accuracy on multiple benchmark datasets for the image classification task, compared to both the hand-designed neural network and the genetic algorithm-based NAS algorithm.