Diffractive Deep Neural Network Identifies Objects at the Speed of Light

Computer vision has changed what we can do with computers, from unlocking your phone by who is in front of it, to making quality checks in manufacturing. These operations can now be accomplished in a fraction of a second, but what if you need it to happen instantly, or at least at the speed of light?

That idea is now a little closer to reality thanks to a “diffractive deep neural network” developed by researchers at the UCLA Samueli School of Engineering. Unlike other vision systems, this network doesn’t sense and image, then process the data using a computer program; instead, it uses diffraction to process light directly.

Light from the sensed object enters a series of polymer layers, which is progressively diffracted and shuttled into a detection array. This array determines how the light network has interpreted the image, allowing this already-interpreted data to be used by more “traditional” computing systems.

As of now, the network has been able to determine handwritten numbers, as well as items of clothing. The physical nature of this device would mean that modifications to how the sensor processes light would take more than changing a few lines of code. However, these layers can be 3D-printed, and the UCLA gadget could be recreated for a mere $50.