Deep neural networks, which are behind many of today’s smart technologies like image recognition or understanding human speech, are getting big and complex. They’re starting to stretch the capabilities of regular computer chips.

Researchers from MIT and other places have developed hardware that uses light instead of electricity for computing. This “photonic” hardware could handle these complex tasks much faster and use less power. The problem was that some of the calculations needed in neural networks couldn’t be done with just light, requiring slow and less efficient electronic components.

Now MIT researchers have built a photonic chip that can do all the necessary calculations for deep neural networks using only light. This chip has a design that allows it to perform both the simple (linear) and more complex (nonlinear) operations of neural networks right on the chip.

The researchers have described the new photonic chip in a paper published in Nature Photonics.

The chip was able to process and classify information in less than half a nanosecond with over 92% accuracy, which is as good as or better than traditional electronic chips.

Faster data processing

This is “an end-to-end system that can run a neural network in optics, at a nanosecond time scale,” says researcher Saumil Bandyopadhyay in an MIT press release.

The chip uses light to encode data, then uses beam splitters to manipulate this light in a way that mimics how neural networks multiply and process data. For the nonlinear operations, they’ve made special devices called Nonlinear Optical Function Units (NOFUs). These devices turn some light into electrical signals for processing, but then quickly turn it back into light.

This new chip was made using the same manufacturing processes as regular computer chips, which means it could be mass-produced easily.

Looking forward, the researchers are thinking about making this chip bigger and connecting it with everyday electronic devices.

They also want to explore new ways to train these optical systems to be even faster and more energy-efficient.

This could lead to breakthroughs in many areas, for example in real-time image processing for cars and devices.

Let us know your thoughts! Sign up for a Mindplex account now, join our Telegram, or follow us on Twitter.