DURIP Electro-Optical Sensor Prototype with Phase/Amplitude Analysis
Abstract
Electro-Optical Sensor Prototype with Phase/Amplitude AnalysisOffice of Naval ResearchTory Cobb Code 321 (Ocean Sensing and SystemsApplications)Approved for public releaseThis DURIP is about building a support workstation to enhance the use of a computing block.The research to build the computing block, called Electro-Optical Deep Learning Accelerators or EODLAs, is funded through the ONR. The workstation envisioned by this grant will allow multiple prototypes to be designed, built, integrated and tested for a variety of stakeholders. The idea behind EODLA is that displays can be arranged in an optical path to enable convolutions of templates with desired signals, pushing the processing heavy convolution-component of modern deep learning off silicon, onto photon-based computation. Our method can work with incoherent light. In this version, we target deep learning algorithms for image processing and computer vision.We want the workstation to allow EODLA development across the board. The ONR grant is about building a computational block. When this is successful, however, there remains a last-mile impact of actually using the device in different applications. These may range from Internet of Things applications for secure building monitoring to underwater or aerial robotics platforms that carry the device.This DURIP is about pushing that into applications, and this will require electronics stations for optics, beamsplitters, already-manufactured micro-scale displays and some electronics. This will enable the use of the prototype in different applications andwill strengthen the chances of this research #leaving the lab# for demonstrations in a variety of diverse scenarios. We are particularly interested in demonstrations in collaboration with Navy labs and research sub-urban environments. A third example application is explosive ordnance remediation and mine countermeasures in surf-zone and terrestrial environments. The research performed in thisgrant would permit multi-modal automated target detection and recognition, visual planning, and control in GPS- and communications-denied environments, and more. This DURIP has three areas of support that allows us to build and deploy EODLAs. The first area of support that we need is fast, efficient retraining to help transfer networks to a form in which it can be put on the EODLA, including binary or low bit networks and the large first layer as described previously. The second area of support is to fine-tune the networkonce it can be put onto the device. This may involve collecting some data from the device and simulating some more in software. Allthis extra data processing and simulation will require additional processing power. This power is requested in this DURIP to enableupdating the networks in near real-time and support the use of EODLA on different platforms. Finally, we may want to capture data, from each layer and update the weights in real time. This will require custom implementation of backprop equations --- a simulation of (say) Pytorch in the optical domain. The inefficiencies that this will cause will create slowness in processing, and this is again why we need dedicated computers to help with the speed up computation of these new weights. With these three areas of support, we will build a workstation with three sub-stations and demonstrate pushing EODLAs out of the lab, onto a variety of platforms, showcasing applications for many different fields.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Nov 08, 2024
- Source ID
- N000142412495
Entities
People
- Sanjeev J. Koppal
Organizations
- Office of Naval Research
- United States Navy
- University of Florida