The article provides a comprehensive overview of optical neural networks (ONNs) and their advancements in recent years. ONNs have emerged as a promising alternative to conventional computing hardware due to their sub-nanosecond latency, low heat dissipation, and high parallelism. The review begins by introducing the design principles and methods of ONNs based on various optical elements, including volume optical components and on-chip components. It then systematically surveys the research progress in non-integrated ONNs and integrated ONNs, highlighting their computational density, nonlinearity, scalability, and practical applications. Non-integrated ONNs, such as those based on 4f systems, diffractive elements, and other bulk optical components, offer reconfigurable functions but are limited by alignment errors and scalability issues. Integrated ONNs, leveraging advanced semiconductor process technologies, provide higher integration levels, better stability, and portability. However, they face challenges in achieving high computing capacity due to thermal crosstalk and power supply requirements. The article also discusses the future development trends and challenges of ONNs, emphasizing the need for further exploration to achieve simultaneous advantages in reconfigurability, nonlinearity, and high computing capacity. The computational density and computing capacity of ONNs are crucial factors, with higher integration levels leading to higher computational densities but potentially lower computing capacities. Overall, ONNs show significant potential in advancing artificial intelligence with novel computing paradigms.The article provides a comprehensive overview of optical neural networks (ONNs) and their advancements in recent years. ONNs have emerged as a promising alternative to conventional computing hardware due to their sub-nanosecond latency, low heat dissipation, and high parallelism. The review begins by introducing the design principles and methods of ONNs based on various optical elements, including volume optical components and on-chip components. It then systematically surveys the research progress in non-integrated ONNs and integrated ONNs, highlighting their computational density, nonlinearity, scalability, and practical applications. Non-integrated ONNs, such as those based on 4f systems, diffractive elements, and other bulk optical components, offer reconfigurable functions but are limited by alignment errors and scalability issues. Integrated ONNs, leveraging advanced semiconductor process technologies, provide higher integration levels, better stability, and portability. However, they face challenges in achieving high computing capacity due to thermal crosstalk and power supply requirements. The article also discusses the future development trends and challenges of ONNs, emphasizing the need for further exploration to achieve simultaneous advantages in reconfigurability, nonlinearity, and high computing capacity. The computational density and computing capacity of ONNs are crucial factors, with higher integration levels leading to higher computational densities but potentially lower computing capacities. Overall, ONNs show significant potential in advancing artificial intelligence with novel computing paradigms.