As an interdisciplinary product of optical (electronic) technology and artificial intelligence technology, optical neural networks can combine both advantages to build a low-power network structure with high computational speed, breaking the bottleneck of traditional electronic neural networks. With the maturity and further development of optical electronic devices manufacturing technology, especially the development of integrated optoelectronic technology, optical neural network technology has made breakthroughs in building feed-forward, recurrent and spiking models with the use of optoelectronic devices. However, compared with the current mature electronic neural networks, optical neural networks still have broad room for improvement in terms of training, integration, expanding scale, and application. On the one hand, the imperfection and low stability of the performance of optoelectronic devices inhibit the training, integration, and expanding scale of the optical neural networks, which puts forward more stringent requirements for building more complex neural network models. On the other hand, the optical neural networks is also limited by the above effects in the application domain, and it is difficult to give full play to the advantages in specific areas. In recent years, although many solutions have been proposed, how to break the bottleneck of the optical neural network at its source requires deep thinking and research. We believe that in the near future, these problems of optical neural networks will be definitely overcome. In addition, people can make better use of the high-speed and low-power advantages brought by optoelectronic technology and artificial intelligence technology to build a greener and more intelligent world.