Optical Neural Network Realized On A Silicon Chip

  • Researchers created a silicon chip that precisely spreads light signal, demonstrating a new neural network design. 
  • The light eliminates interference caused used by electrical charge, and can travel faster and farther. 

Developing energy-efficient and highly compact interconnects has been a key research goal for integrated photonics. They have a wide range of applications, including effective telecommunications and high-bandwidth cross-chip communications in CMOS devices.

Many scientists around the world are working on artificial neural network circuits to emulate the human brain. However, the traditional electrical wiring of semiconductor circuits isn’t capable of handling the intensely complex routing needed for advanced neural networks.

Recently, scientists at the National Institute of Standards and Technology developed a silicon chip that precisely diffuses optical signals across a tiny brain-like grid, demonstrating a new neural network design.

The artificial neural networks have shown exceptional capabilities in learning and modeling non-linear complex problems, including image processing, character recognition, and data forecasting. Now, the research team has used light signals (rather than electric signals) to implement these neural networks.

Advantages of Using Light Over Electric Signals

The main reason for using light instead of electric signals is light eliminates interference caused used by electrical charge, and thus can achieve longer communication with higher speed and lower power.

It can enhance the performance for scientific data analysis. This includes investigation of quantum data science, searches for exoplanets and development of autonomous vehicle control systems.

A traditional computer processes data via coded rules or algorithms, while the neural network depends on multiple connections among processing units called neurons. The multiple layers of neurons can be trained to perform some specific tasks. Typically, a neuromorphic machine contains a large, complicated structure of neural networks.

How Did They Build An Optical Chip?

The new silicon chip uses light signals by stacking (vertically) 2 photonic waveguide layers. This confines light into narrower lines to route light signals. More specifically, the stacking of waveguides enables dense integration with low-cross talk and low-loss waveguide crossings.

Reference: APL Photonics | doi:10.1063/1.5039641 | NIST

The 3D design allows for complex routing schemes, and it can be integrated with additional layers to perform more complicated tasks.

In this work, they presented stacked waveguides that create a 3D grid with 10 inputs each connecting to 10 outputs. Basically, it’s a routing between 2 layers of a feed-forward neural network with a total of 100 receivers.

Optical Neural Network on silicon chipPhotonic routing manifold | Credit: Chiles / NIST

They used silicon nitride to build these waveguides (each is 400 nanometers thick and 800 nanometers wide) and fabricated them on a silicon wafer. They also developed a dedicated program to automatically produce signal routing, with appropriate (configurable) connectivity level between the neurons.

Then, they used an optical fiber to direct a laser light into the silicon chip. The aim was to route every single input to all outputs, following a distribution pattern for light power or intensity. Different levels of power show a different degree of connectivity and pattern within the circuit.

The researchers showed 2 schemes to control output intensity –

  1. Uniform: all outputs receive the same power.
  2. Bell Curve distribution: most of the power is transferred to middle neurons.

Read: A New Form Of Light That Could Make Quantum Computing Possible With Photons

To precisely analyze the outcomes, they created images of the signals coming out of the last layer. The output had low error rates and precise power distribution. At 1320 nanometers of wavelength, the uniform and bell curve distribution were found to have mean output power errors of 0,7 and 0,9 dB.

Written by
Varun Kumar

Varun Kumar is a professional science and technology journalist and a big fan of AI, machines, and space exploration. He received a Master's degree in computer science from Indraprastha University. To find out about his latest projects, feel free to directly email him at [email protected] 

View all articles
Leave a reply