Deep Learning has become increasingly popular since 2006. It has an outstanding capability to extract and represent the features of raw data and it has been applied to many domains, such as image processing, pattern recognition, computer vision, machine translation, natural language processing, and autopilot. While the advantages of deep learning methods are widely accepted, the limitations are not well studied. This thesis studies cases where deep learning methods lose their advantages over traditional methods. Our experiments show that, when the neighbouring proximity disappears, deep learning methods are significantly less powerful than traditional methods. Our work not only clearly indicates that deep structure methods cannot fully replace traditional shallow methods but also shows the potential risks of applying deep learning to autopilot.
