商空间粒变换的深度特征表示

Deep Feature Representation Based on Granular Transaction of Quotient Space Theory

目前,大数据问题亟待解决,关键就是对问题的特征描述.目前特征描述最流行的理论是深度学习理论,但深层结构共需要多少层,每层需要多少特征?这是深度学习最需要解决的问题.引入商空间理论对深度学习理论进行改进,根据粒度变换原理对问题特征进行深层表示,克服深度学习理论中深度不确定,特征描述不明确的缺点.首先根据商空间理论的粒度变换原则,在多粒度空间分层描述问题特征,从而形成多层的深度特征表示.接着,根据商空间粒度变换的描述特性,在不同粒度空间对问题进行求解.最后,作者选取Letter-recognition数据集进行实验,实验结果表明本文所提的深度特征表示法可以自动将问题分为多层结构,分层描述问题的特征,提升了问题求解精度.

At present, big data is a matter of urgent doubt. The key of this issue is feature representation of problems. The most popular theory of feature representation is Deep Learning. But how much layers are need？ How much features are need in each layer？ These are our most urgent problem in Deep Learning. In this paper, author introduces Quotient Space Theory { QST ） to improve Deep Learn- ing Theory. The feature is represented automatically by deep layers. So the uncertainty of number of deep layers is overcome, and the fault is also conquered that feature representation is indefinite. Author hierarchically descries features of problem in multi-granular spaces to form multilayer feature representation. The problem is solved in different granular space based on the granular transaction principle of QST. The experiment on Letter-recognition data set shows that the deep feature representation proposed in this paper ex- presses the whole problem with hierarchical structure by itself. The feature is described hierarchically and the solution precision is raised.