Neighbouring Proximity - An Key Impact Factor of Deep Machine Learning

Abstract

Deep Learning has become increasingly popular since Alexnet was proposed. It has been applied to many domains such as pattern recognition, computer vision, machine translation, and natural language processing. While the advantages of deep learning methods are wildly accepted, the limitations of them are not well researched. In this paper, we present our study and analysis of cases where deep learning methods lose their advantages over traditional methods. Our experiments show that, when the neighbouring proximity disappears, the accuracy of deep learning methods is at most as good as, if not worse than, that of advanced traditional shallow methods. As the resources that traditional shallow methods needed are always much less than deep learning methods. We conclude that, in situations where neighbouring elements of input samples do not have proximity, deep learning methods are significantly less powerful than traditional methods. Furthermore, this clearly indicates that deep structure methods cannot fully replace traditional shallow methods.