ACTIVE BAYESIAN DEEP LEARNING WITH AN ACOUSTIC VECTOR SENSOR

Loading...
Thumbnail Image
Authors
Atchley, Sabrina L.
Subjects
Bayesian neural networks
active learning
Advisors
Orescanin, Marko
Date of Issue
2021-09
Date
Publisher
Monterey, CA; Naval Postgraduate School
Language
Abstract
Traditional passive monitoring of the ocean’s acoustic signals is conducted with an omnidirectional hydrophone, which only measures acoustic pressure. Vector sensors, unlike hydrophones, respond to both the acoustic pressure and the vector motion of water, providing additional information. This thesis focuses on utilizing vector sensor data as input to a neural network and studies the advantage of utilizing all four channels over single-channel data from the acoustic pressure sensor. A Bayesian deep learning approach is used to build multi-class classification models that provide estimates of uncertainty. The best model had an F1 score of .798 using single-channel data and .81 when using four-channel data from the vector sensor. However, the addition of information from the four-channel signal significantly reduces predictive uncertainty, demonstrating the advantage of utilizing all four channels for passive sonar classification. Next, active learning is examined, an algorithm that typically depends on uncertainty estimates to select the best training data. This is likely the first study on active learning with Bayesian deep learning models in passive sonar classification. With active learning using 23% of the training data, we trained within two percent of the F1 score compared to the entire training data. Additionally, the active learning experiments demonstrated that uncertainty-based acquisition functions increased performance using four channels over single-channel data.
Type
Thesis
Description
Organization
Identifiers
NPS Report Number
Sponsors
Funder
Format
75 p.
Citation
Distribution Statement
Distribution Statement A. Approved for public release: Distribution is unlimited.
Rights
This publication is a work of the U.S. Government as defined in Title 17, United States Code, Section 101. Copyright protection is not available for this work in the United States.
Collections