BYU

Abstract by David Van Komen

Personal Infomation


Presenter's Name

David Van Komen

Degree Level

Masters

Abstract Infomation


Department

Physics and Astronomy

Faculty Advisor

Tracianne Neilsen

Title

Using multi-task deep learning to predict range and seabed type in the ocean

Abstract

Localization of sources and predicting the seabed type in the ocean are important problems in underwater acoustics. The way sound travels in the ocean is dependent on many different factors, so using data-driven deep learning techniques to avoid computationally expensive model-based solutions is desirable. Deep learning solutions have shown potential in solving these problems, though an important question to consider is how the predictions should be made. A deep learning model can be configured to classify or regress predictions or configured to use some combination when making simultaneous predictions (multi-task). This question is explored by comparing predictions of measured underwater explosions from networks trained on simulated explosions. Instead of just regressing or classifying both predictions, classifying the seabed and regressing the seabed type simultaneously gives better results. The results of this experiment illustrate the need to use the proper type of network outputs.