Diffractive Optical Networks Utilize Object Shifts for Performance Boost

Optical computing has gained significant interest in machine learning applications due to optics’ massive parallelism and bandwidth. Diffractive networks supply one such computing paradigm based upon the transformation of the input light as it diffracts via a set of spatially-engineered surfaces, performing computation at the speed of light propagation without needing any external power apart from the input light beam. Among many other applications, diffractive networks have been shown to perform all-optical classification of input objects.

Scientists at the University of California, Los Angeles (UCLA), led by Professor Aydogan Ozcan, introduced a “time-lapse” scheme to considerably improve the image classification accuracy of diffractive optical networks on complex input objects. The discoveries are released in the journal Advanced Intelligent Systems.

In this scheme, the thing and/or the diffractive network are moved relative to each other during the exposure of the output detectors. Such a “time-lapse” scheme has formerly been utilized to achieve super-resolution imaging, for instance, in security cameras, by capturing numerous pictures of a scene with lateral camera movements.

Drawing inspiration from the success of time-lapse super-resolution imaging, UCLA scientists have used “time-lapse diffractive networks” to achieve over 62% blind testing accuracy on the all-optical classification of CIFAR-10 pictures, a publicly available dataset containing pictures of airplanes, cars, cats, and so on. Their results achieved a significant boost over the time-static diffractive optical networks.

Exploring optical computing

The same research study group had formerly demonstrated ensemble learning of diffractive networks, where numerous diffractive networks operated in unison to improve photo classification accuracy. Nevertheless, with the incorporation of the time-lapse scheme, it is conceivable to outdo an ensemble of over 15 networks with a single standalone diffractive network, considerably reducing the footprint of the diffractive system while removing the complexities of physical alignment and synchronization of several individual networks. The scientists additionally explored the incorporation of ensemble knowing into time-lapse picture classification, which revealed higher than 65% blind testing precision in classifying CIFAR-10 pictures.

For the physical implementation of the time-lapse classification scheme presented, the most straightforward method would exploit the natural vibration of the objects or the diffractive camera during imaging and enable the benefit of time-lapse to be collected with no additional cost apart from a slight boost in the inference time because of detector signal integration during the jitter.

This study on time-lapse image classification is a demonstration of using the temporal degrees of freedom of optical fields for optical computing. It presents a huge step forward toward all-optical spatiotemporal information processing with compact, affordable, and passive products.

Read the original article on TechXplore.

Read more: A Newly Discovered Form of Salty Ice Could Exist on the Surface of Extraterrestrial Moons