Fiber non-linearity remains a major limiting factor in long distance transmission systems today.

Researchers at Princeton Lightwave Lab and NEC Laboratory America have created a real-time neural network on an integrated photonic chip using silicon photonics. This technology can be useful for trans-Pacific power lines up to 10,000 km long and can also help overcome the adverse effects of non-linear fiber. This is a great example of how photonics is superior to electronics in AI applications.

Existing fixed communication networks, wireless infrastructure and data centers are heavily dependent on optical communication systems that transmit information over optical channels. In the last decade, the growth of the Internet has been largely supported by a technology called digital signal processing (DSP), which can reduce transmission distortion. However, DSP is implemented using CMOS integrated circuits and has reached its limits in terms of power dissipation, density, and Moore’s Law engineering.

Therefore, the distortion caused by fiber non-linearity cannot be compensated by the DSP, as this would require too much computing power and resources. Thus, fiber non-linearity remains a major limiting factor in long distance transmission systems.

Chaoran Huang, a researcher at NEC Laboratories America, Inc., and colleagues have developed a photonic neural network based on high-performance waveguides and photonic devices, including photodetectors and modulators, originally designed for optical communication.

The optical modulator converts the electrical photocurrent into optical energy using the photocurrent generated during this initial process. As a result, optical modulators serve as artificial neurons in the photonic network.

In addition, the silicon neural network created by the researchers is programmable and based on the so-called broadcast and weight protocol. This architecture uses neurons that are multiplexed into a waveguide to create optical signals of a specific wavelength that are transmitted to all other neurons.

Silicon photonic-electronic neural networks have clear advantages over electronic neuromorphic circuits in terms of power dissipation, latency, crosstalk, and throughput. These specific properties make silicon photonic neural networks ideal for building large systems with many artificial neurons on a single chip using only a few interconnect waveguides.

This technique can come in handy in machine learning, non-linear programming, and signal processing.