On the Construction of Hierarchical Broad Learning Neural Network: An Alternative Way of Deep Learning

摘要

In this paper, we proposed an alternative way of deep learning, named as Hierarchical Broad Learning (HBL) neural network which forms a neural network with three layers. It is based on Broad Learning System (BLS), hence HBL inherits the characteristics of feature and enhancement nodes in a neural network. In HBL, parameters of the hidden layer are trained in a forward manner so that the weights of the current layer can be fixed without fine-tuning once the input layer of feature and enhancement node is established. Due to the multilayer architecture of HBL, it can handle image or video data in the more effective way. Experimentation of the proposed HBL is performed on two benchmark datasets, MNIST and NYU NORB (object recognition dataset). The results show that the training time of proposed HBL is effectively less than the existing state of the art learning methods. It also achieves better accuracy and generalization performance than the Broad Learning System.