Artificial Neural Network for Optical Chips

A neural network can be trained using an optical circuit. In the full network there would be several of these linked together. The laser inputs (green) encode information that is carried through the chip by optical waveguides. (Source: T. W. Hughes, Stanford Univ.)

Researchers have shown that it is possible to train arti­ficial neural networks directly on an optical chip. The signi­ficant break­through demon­strates that an optical circuit can perform a critical function of an elec­tronics-based arti­ficial neural network and could lead to less expensive, faster and more energy effi­cient ways to perform complex tasks such as speech or image recog­nition.

“Using an optical chip to perform neural network compu­tations more efficiently than is possible with digital computers could allow more complex problems to be solved,” said research team leader Shanhui Fan of Stanford Univer­sity. “This would enhance the capa­bility of arti­ficial neural networks to perform tasks required for self-driving cars or to formulate an appro­priate response to a spoken question, for example. It could also improve our lives in ways we can’t imagine now.”

An artificial neural network is a type of artificial intelli­gence that uses connected units to process infor­mation in a manner similar to the way the brain processes infor­mation. Using these networks to perform a complex task, for instance voice recog­nition, requires the critical step of training the algorithms to cate­gorize inputs, such as different words. Although optical arti­ficial neural networks were recently demonstrated experi­mentally, the training step was performed using a model on a tradi­tional digital computer and the final settings were then imported into the optical circuit. Now, the researchers report a method for training these networks directly in the device by imple­menting an optical analogue of the ‘back­propagation’ algorithm, which is the standard way to train conven­tional neural networks.

“Using a physical device rather than a computer model for training makes the process more accurate,” said Tyler W. Hughes. “Also, because the training step is a very computa­tionally expensive part of the implemen­tation of the neural network, performing this step optically is key to improving the computa­tional effi­ciency, speed and power consump­tion of arti­ficial networks.” Although neural network processing is typi­cally performed using a tradi­tional computer, there are signi­ficant efforts to design hardware opti­mized speci­fically for neural network computing. Optics-based devices are of great interest because they can perform compu­tations in parallel while using less energy than electronic devices.

In the new work, the researchers overcame a signi­ficant challenge to imple­menting an all-optical neural network by designing an optical chip that replicates the way that conven­tional computers train neural networks. An arti­ficial neural network can be thought of as a black box with a number of knobs. During the training step, these knobs are each turned a little and then the system is tested to see if the perfor­mance of the algo­rithms improved. “Our method not only helps predict which direc­tion to turn the knobs but also how much you should turn each knob to get you closer to the desired perfor­mance,” said Hughes. “Our approach speeds up training signi­ficantly, especially for large networks, because we get infor­mation about each knob in parallel.”

The new training protocol operates on optical circuits with tunable beam splitters that are adjusted by changing the settings of optical phase shifters. Laser beams encoding infor­mation to be processed are fired into the optical circuit and carried by optical wave­guides through the beam splitters, which are adjusted like knobs to train the neural network algo­rithms.

In the new training protocol, the laser is first fed through the optical circuit. Upon exiting the device, the dif­ference from the expected outcome is calcu­lated. This information is then used to generate a new light signal, which is sent back through the optical network in the opposite direc­tion. By measuring the optical intensity around each beam splitter during this process, the researchers showed how to detect, in parallel, how the neural network perfor­mance will change with respect to each beam splitter’s setting. The phase shifter settings can be changed based on this infor­mation, and the process may be repeated until the neural network produces the desired outcome.

The researchers tested their training technique with optical simu­lations by teaching an algo­rithm to perform compli­cated functions, such as picking out complex features within a set of points. They found that the optical imple­mentation performed similarly to a conven­tional computer. “Our work demon­strates that you can use the laws of physics to implement computer science algo­rithms,” said Fan. “By training these networks in the optical domain, it shows that optical neural network systems could be built to carry out certain func­tionalities using optics alone.”

The researchers plan to further optimize the system and want to use it to implement a practical appli­cation of a neural network task. The general approach they designed could be used with various neural network archi­tectures and for other appli­cations such as recon­figurable optics. (Source: OSA)

Reference: T. W. Hughes et al.: Training of photonic neural networks through in situ backpropagation and gradient measurement, Optica 5, 864 (2018); DOI:

Link: Ginzton Laboratory, Dept. of Electrical Engineering, Stanford University, Stanford, USA

Speak Your Mind