On November 19, 2016 (October 20, 2016 in the lunar calendar), the world's first photonic neural network was born: 1960 times faster than traditional methods. Artificial intelligence technologies represented by machine learning neural networks are changing many aspects of our lives, and the demand for the data processing capabilities needed to support these technologies is growing. The development of neuromorphic chips for the structure of neural networks is expected to greatly improve the performance of neural networks. Recently, scientists such as Alexander N.Tait from Princeton University published a paper titled "Neuromorphic Silicon Photonics" on arXiv, introducing the world's first "Photonic Neural Network." Compared with our current mainstream electron-based processing methods, the photonic method can achieve faster speeds and achieve higher bandwidth. According to a report by MITTechnology Review, the core of the problem of developing photonic neural networks is to produce a device where each node has the same response characteristics to serve as neurons. These nodes take the form of miniature ring waveguides that are etched into a silicon substrate through which light can circulate. When this light is released and the output of a laser working at the threshold is modulated, a slight change in the input light in the environment can have a huge impact on the output of the laser. The key is that each node in the system operates at a certain optical wavelength-a technique known as wavelength division multiplexing. Light from each node is detected and summed by total power before being fed into the laser. The laser output is then fed back to the node to create a feedback loop with nonlinear characteristics. So to what extent does this non-linearity simulate neural behavior? Research has shown that its output is mathematically equivalent to the output of a type called a continuous-time recurrent neural network. The researchers developed a 49-node silicon photonic neural network for proof of concept-experiments showed to be 1960 times faster than traditional methods in an experimental differential system simulation mission! Of course, confirmatory experiments may not be suitable for actual application scenarios, but there is no doubt that this research provides an important impetus for the development of photon-based neural networks. This is expected to provide us with the solution we need today when bandwidth and speed demand are increasing. Paper: Neuromorphic Silicon Photonics We report the first observation of an integrated analog photonic network, in which connections are configured through micronizing weightbanks and electro-optic modulators used for the first time as photonic neurons. This mathematical isomorphism between the silicon photonic circuit and the continuous neural model is demonstrated through dynamic bifurcation analysis. Existing neural engineering tools can take advantage of this isomorphism to adapt to silicon photonic information processing systems. We used a "neural compiler" to program a 49-node silicon photonic neural network that simulates traditional neural networks and is expected to perform 1960 times better than traditional methods in an experimental differential system simulation mission. Photonic neural networks that utilize silicon photonics platforms can access ultra-fast information processing environments for radio, control and scientific computing. Figure 1: STAR broadcast-and-weight network with modulators used as neurons. MRR: microresonator, BPD: balanced photodiode, LD: laser diode, MZM: Mach-Zehnder modulator, WDM: wavelength-division multiplexer. Figure 2: Experimental configuration with 2 MZM neurons and an external input, wavelength multiplexed in an arrayed waveguide grating (AWG) and coupled to an on-chip broadcast-weight network. This 2×2 cyclic network is configured by MRR weights, w11, w12, etc. The neuron state is represented by voltages s1 and s2 of low-pass filtered transimpedance amplifiers.