深度神经网络的快速学习算法

A Fast Learning Algorithm for Deep Neural Network

通过分析传统的多层感知器和反向传播算法的不足,设计了一个全新的网络结构SC-MLP和提出了与之对应的全新的学习算法NBP,主要是实现权值的模和固定,这样可以加速训练的速度.在高维数据分类的实证分析中,以手写数字数据库为例,构建了一个深度神经网络,并对比各种训练算法.实验表明,NBP学习算法对于深度神经网络具有良好的学习效果,明显优于传统的反向传播算法,并且在精度上与深度学习算法相当,但是速度快.

By analyzing the deficiencies of the traditional multilayer perceptrons and back - propagation algorithm, this paper designed a new network architecture SC - MLP and proposed the corresponding new learning algorithm NBP, mainly realizing the fixed norm of weights, so we can accelerate training speed. In the empirical analysis of high- dimensional data classification, for example handwritten digital databases, we construct a deep neural network and compare the various training algorithms. The results show that the NBP learning algorithm has good learning effect for learning neural network, which is faster but much better than the traditional back - propagation algorithm, and as good as the “deep learning” algorithms in the precision.