Photonic Convolutional Processor for Network Edge-Computing
Abstract
ONR: Special Notice N00014-19-R-S002Photonic Convolutional Processor for Network Edge-ComputingPI: Volker Sorger (GWU); Team: UT Austin, UCLAABSTRACTThe current need to push data processing and computing power away from centralized points of large networks to locations closer to the user requires new paradigms and architectures able to efficiently handle vast data at high-speed in network-centered warfare. Moreover, cognitive tasks aiming to extract precious features from such large data-set via convolutional neural networks (CNN) are computational intensive but are limited by intrinsic properties of CMOS electronics. Therefore, simultaneously bypassing sending data across long routes to data centers or cloud and suppressing the wire-charging limiting electronics is therefore necessary for sensitively reducing latency and enhancing computation efficiency.This project develops two versions of a photonic convolutional processors based on the optical 4f approach and are capable to speed-up Fourier-optical convolution processing by several orders of magnitude. Generation 1 of this system explores a 4f free-space convolution processor based on Digital micromirror devices capable of modulating a low power coherent light at high resolution (2Mpx) and moderate speed (1kHz for 8bit), which replace slow-rate (tens of Hz) spatial light modulators (SLM) that are based on liquid crystal technology. This processor, able to handle vast data at moderate rate, could be employed for fast image filtering applications and perform tasks on offline trained CNN, that require inference over large images and could be particularly beneficial for applications such as pattern recognition and super resolution in geolocalization.The implementation and characterization of the free-space convolutional processor serving as an explorative study paves the way for a second-generation (Gen-2) CNN processor leveraging Silicon-based photonic integrated circuits forming a joint transform correlator-based approach. Thisenhanced system exploits GHz-fast data handling; data loading front end, convolutional filtering, and back-end data registration is performed at nanosecond short delays, unlike the kHz-slow implementation using SLMs. Conceptually, the data is kept ~longer~ in the optical domain in this photonic co-processor, thus avoid electro-optic conversion where able, or speed it up usingintegrated photonics with 10 s GHz fast devices available through silicon photonic foundries such as AIM. This project is performed by a team comprised of experts from both industry and academia featuring mutual expertise with high project relevance in device, circuit, and system-level R&D including prototyping in opto-electronics, integrated photonics, RF circuits, computer engineering and architecture, neural networks, and data-analytics.Since the Gen-2 processor performs convolutions at ultra-high throughput, applications requiring demanding computation and real-time execution will be beneficiaries, thus adding new capabilities in electronic warfare support to include enhanced situational awareness and threat warning. Such tactical sensing for real-time includes artificial intelligence using neural networks enabling in applications such as synthetic aperture radar for automated target recognition, classification of electromagnetic signatures, signal intelligence and image processing in hyper-spectral filtering - all key applications for DOD. If successful, this project will deliver prototypes capable of surpassing current 4f systems by a factor of several orders of magnitude in terms of response time-throughput product and thus pave the way for field-testing.Approved for Public Release
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 20, 2019
- Source ID
- N000141912595
Entities
People
- Volker Sorger
Organizations
- George Washington University
- Office of Naval Research
- United States Navy