Advancing Geo-localization with High-Performance-Computing
Abstract
Geolocalization is the process an agent uses to determine where it ison earth using whatever local data is available. It is not always possible for an agent to receive GPS signal #the agent might be underwater, where GPS signals do not propagate, or is in a GPS-denied area. Recentgroundbreaking research on polarization based geolocalization at University of Illinois Urbana-Champaignhas shown that an underwater agent, simply by looking at the underwater polarization patterns, can tellwhere it is on earth to astonishing accuracy. This framework works even in turbid (i.e., low visibility)waters or under moonlight.The key to accurately predict geolocationfrom underwater polarization images at night or day istraining of moderately sized neural network based machine learning models with immense quantities ofdata produced by polarization cameras at many locations on earth. To train, one must pass as much dataas possible through the relevant machine learning model. The accuracy improvements of a data drivenmethod compared to a parametric model approach are staggering: parametric model based geolocalizationerror is over 1,500km and around 20 km with a neural network approach. It is reasonable to believe thatwith more data produced from novel bioinspired multi spectral and polarization sensitive cameras andwithcarefully crafted network architectures, the geolocalization accuracy can be significantly improved.To achieve this DoD relevant goal, massive data training throughput and a careful search of architecturesare required. Similar demands are observed across manyDoD-funded, NSF-funded and NIH-funded projectson our campus: large data sets that are generated or available for various research projects require significantcomputational resources to find an optimal machine learning n etwork. This computationally dauntingtask requires significant number of graphical processors, fast data storage and fast networking capabilities,which are currently lacking on the University of Illinois campus.To meet the computational challenges associated with advanced machine learning research, we proposeto purchase a high performance computing instrument to support DoD-funded (specifically researchfunded by ONR, DARPA and ARL) andnon DoD-funded research programs at the University of Illinois.The proposed instrument will provide state of the art computational resources for advanced and rapiddevelopment of machine learning algorithms for at least seven research groups across three engineeringdepartments at the University of Illinois. Furthermore, these computational resources will be integrated inthe machine learning curriculum taught across all departments in the school of engineering and providetraining for the next generation of engineers and data scientists. Specifically, we will use the equipment forclass projects that use publicly available DoD-related project data sets which we will use to spur even moreinterest in DoD research.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 12, 2023
- Source ID
- N000142312166
Entities
People
- Viktor Gruev
Organizations
- Office of Naval Research
- United States Navy
- University of Illinois Urbana–Champaign