Procedural Neural Network Based on Statistical Features

This paper deals with a novel model which called procedural neural network model based on some statistical features of temporal data set. As large amount of information included in spatio-temporal data problem, computational complexity is a key issue for procedural neural networks. Some statistical features, such as expectation, variation, play important roles in expression a series of variable data with respect to time. So these statistical features are introduced as aggregation mapping from a temporal domain to vector space in procedural neural networks instead of calculating each input data on temporal axes. By this strategy, computational complexity in the procedural neural networks is reduced down deeply as that of the traditional static neural networks. Also learning algorithm for this kind of procedural neural network is proposed and a stock price prediction problem is given as a test example for this model.