As a three-dimensional (3D) display method, computer-generated holography (CGH) can achieve accurate reconstructions of the target light fields based on diffractive optics. It has broad applications in the metaverse, head-mounted display, head-up display, etc. High-speed calculation and high-quality reconstruction of phase-only holograms (POHs) are key issues that should be emphasized in this field. In recent years, the leapfrog development of deep learning has provided a novel path to address this challenge. In this review, the basic principles and classifications of CGH are briefly introduced. Then, the existing CGH methods based on deep learning are summarized. The advantages and disadvantages of various methods are compared. Finally, the possible research directions and challenges of this field are prospected.