Data Driven Methods for Structure Learning in Underwater Acoustic Modeling
Abstract
The proposed e ort consists of four complementary tasks:Task 1: Navy systems are increasingly limited by the data and computationa"l resources required to make use of traditional ocean physical models, implying that higher levels of performance will require models that capture ner-grain phenomena and characteristics in the ocean physics presently unmodeled, across a wider range of spatial and temporal scales. Accordingly, this task focuses on the development of such re ned multiscale physical modelingof ocean acoustic processes, including propagation and (ambient) noise phenomena. Task 2: Existing models for ocean acoustic propagation are often characterized by large numbers of parameters, despite application performance being dominated by comparatively few e ective degrees of freedom. Accordingly, this task focuses on the development of nonlinear dimension reduction (NLDR) techniques for spatio-temporal oce"an acoustic data to learn the underlying low-dimensional ocean acoustic manifold."""" Task 3: Typically, inference is carried out on" low-dimensional features extracted fromunderwater acoustic signals. The design of such feature extraction has been guided by a combination of underlying physical models and heuristics, in an unscalable manner. However,a variety of powerful learning structures have emerged, including the increasingly popular deep neural network architectures, that automate the process of selecting features for a target inference task based on data, resulting in scalable, high-performance solutions. Accordingly this task focuses on the development of such learning and inference architectures, incorporating constraints and structure speci c to acoustic signals in ocean environments.Task 4: Traditional underwater acoustic physical models are invariably idealized in a variety of ways to facilitate their analysis and application. While modern learning methods can produce models directly from data that capture key neglected nonidealities, the amount of data required render these techniques impractical. In this task, we will address this by developing a transfer learning methodology in which synthetic data generated by a simulator governed by the idealized physics is used in conjunction with a comparatively small amount ofreal data to augment the idealized physics in a data-driven manner, producing an augemented model" that can be directly applied to a diversity of tasks of Navy interest.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 30, 2019
- Source ID
- N000141912665
Entities
People
- Gregory Wayne Wornell
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy