A Highly-Parallel Optical Processor, Designed Through Deep Learning, Could Accelerate Deep Learning

A team of researchers from the University of California have developed an optical processor designed to offer massive parallelization for deep-learning workloads — by giving a deep-learning workload the job of designing it.

"Large-scale linear operations are the cornerstone for performing complex computational tasks. Using optical computing to perform linear transformations offers potential advantages in terms of speed, parallelism, and scalability," the researchers explain of the reason for looking outside traditional electronics. "Previously, the design of successive spatially engineered diffractive surfaces forming an optical network was demonstrated to perform statistical inference and compute an arbitrary complex-valued linear transformation using narrowband illumination."

To get the performance required, however, the team needed to design a massively parallel broadband diffractive neural network for all-optical computation — and, in an example of "set a thief to catch a thief" logic, put the design of the required broadband diffractive processor in the hands of the very type of machine learning system it's designed to accelerate.

The resulting processor is made up of multiple diffractive layers, made up of passive materials that allow the transmission of light. Information is encoded as wavelengths of light and inserted into the processor, with each particular wavelength dedicated to a particular function or linear transformation.

"These target transformations can be specifically assigned for distinct functions such as image classification and segmentation," explains project lead Aydogan Ozcan, professor at the University of California's Samueli School of Engineering, "or they can be dedicated to computing different convolutional filter operations or fully connected layers in a neural network. All these linear transforms or desired functions are executed simultaneously at the speed of light, where each desired function is assigned to a unique wavelength. This allows the broadband optical processor to compute with extreme throughput and parallelism."

"Such massively parallel, wavelength-multiplexed diffractive processors will be useful for designing high-throughput intelligent machine vision systems and hyperspectral processors," Ozcan predicts, "and could inspire numerous applications across various fields, including biomedical imaging, remote sensing, analytical chemistry, and material science."

The team's work has been published in the journal Advanced Photonics with an open-access copy available on the SPIE Digital Library.