【摘要】 提高神经网络(NN)的收敛速率和预测精度一直是人工智能领域的一个挑战性问题,尽管许多研究人员已在研究中使用小批量数据训练神经网络获得了较好的效果,但是这些方法并不够灵活.针对这个问题,该文提出了一种新的数据预处理算法即Fibonacci采样算法.根据Fibonacci数列规则绘制一个新的训练数据序列,这不仅可以恢复小批量数据的划分,而且还可以提供更灵活的批量规模的选择.实验结果表明:在梯度下降之前,Fibonacci数据划分序列能得到较好的实验结果.在应用于相同的单个CNN(5层卷积神经网络)和AlexNet中,Fibonacci采样算法比传统的小批量梯度下降算法能获得更高的准确度和更低的损失值,并且在几种通用网络(LeNet、AlexNet、VGG-16、GoogLeNet)上的性能也取得显著提升.更多还原

【Abstract】 Improving the convergence speed and prediction accuracy of neural network(NN) has always been a challenging problem in the field of computer artificial intelligence.Although many researchers have studied using small batches of data to train neural networks to obtain better results, these method are not enough flexible.Aiming at this problem, the Fibonacci sampling algorithm that is a new data preprocessing algorithm is proposed.A new training data sequence is drawn according to the Fibonacci sequence rules, which can not only recover the small batch data partition, but also provide more flexible batch size selection.Experiments show that before gradient descent, the Fibonacci data partition sequence can get good experimental results.In the same single CNN(five-layer convolutional neural network) and AlexNet, the Fibonacci sampling algorithm can obtain higher accuracy and lower loss than traditional mini-batch gradient descent algorithm, and it can be used in several general networks such as LeNet, AlexNet, VGG-16,GoogLeNet.更多还原