New technique uses light to train artificial neural networks

Optical chips can be trained to support networks that enable artificial intelligence, with significant implications for machine learning in imaging.

A new study, published in Optica, contends that artificial neural networks can be recorded directly on an optical chip, potentially avoiding the need to use conventional computers for machine learning.

Artificial neural networks are one of the most important algorithms in artificial intelligence and machine learning. They must be trained to perform a given task, which is done by fine-tuning the algorithm and testing it until the desired function is achieved.

Currently, training is done “off-chip” on traditional computers, and final weights are imported into a physical device.

The researchers, from Stanford University, wanted to determine if optics and photonics, technologies that generate and harness light, could be used to improve the training of these algorithms.

They created an optical chip that replicates the way that conventional computers train neural networks. Their 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.

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 can be repeated until the algorithm produces the desired outcome.

The researchers tested their training technique with optical simulations by teaching an algorithm to perform difficult tasks, such as picking out complex features within a set of points. They found that the optical implementation performed similarly to that of a conventional computer.

“Our work should enhance the appeal of photonic circuits in deep learning applications, allowing for training to happen directly inside the device in an efficient and scalable manner,” the study authors conclude.

The optical hardware implementation of these neural networks is a “promising alternative” to the use of conventional computers, according to Tyler W. Hughes, with the department of applied physics at Stanford and first author of the paper

“Using a physical device rather than a computer model for training makes the process more accurate. 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,” he says.

The new methodology could have a major impact on imaging.

“Medical imaging has been one of the most successful applications of machine learning in recent years, and I have seen countless interesting demonstrations of using neural networks in this field, for example for tumor detection, diagnosis, or non-invasive imaging,” says Hughes.

“In the medical imaging field, one could imagine the training stage as where the network would learn whether a tumor is malignant or benign, for example. We present an efficient way to perform this directly on the optical chip and our finding could lead to more powerful machine learning-based imaging techniques being used in the future,” he points out.