Domain Transfer for Continuity of Performance Across Synthetic Aperture Sonar Systems
Abstract
Using novel techniques for formalizing mathematical specifications of complex systems, we will develop a method -for representing high-level machine learning tasks that can be used to describe the problems -facing naval mine counter-measure (MCM) research and development teams. These specifications will represent -machine learning systems at a higher level of abstraction than layers and neurons, which should enable the -automatic construction of training algorithms. We will study these techniques by exploring loss function -selection and training algorithm generation. These approaches will be evaluated on synthetic aperture sonar data -sets in the context of MCM concepts of operation. -This goal will be achieved through investigation of the following research tasks: -(1) The software can emit specifications of machine learning models that can be trained with existing ML -libraries. -(2) The software can generate a list of pretraining strategies and execute a search over loss functions to rank -alternative problem formulations. -(3) The software can train the continuity of performance diagram morphism shown in Fig. 3. -(4) The continuity of performance training in release 3 has been evaluated on SAS Data. -(5) The software has improvements to the diagram morphism training methods in release 1 and 2. -(6) These improvements have been evaluated on SAS data.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 12, 2023
- Source ID
- N000142312339
Entities
People
- James Fairbanks
Organizations
- Office of Naval Research
- United States Navy
- University of Florida