Machine learning in acoustic applications with wave physics models, a tutorial

Abstract

Through recent decades of intensive research, machine learning has established its great potential in areas such as social networking, e-commerce, computer vision, natural language processing, and robotics. Recently, it has also started to capture the attention of the acoustic research community where the problems have wave physics basis and the research has focused on related areas such as signal/image processing. This is evident in the significantly increased number of related papers presented at recent ASA meetings. The challenge appears to concerning several aspects: i) What types of acoustic applications can be formulated as ML problems in a way that could provide interesting results and potentially better performance? ii) What would be the relationship between wave physics models and machine learning methods? iii) What is the dataset requirement of machine learning that is potentially different from say, signal processing techniques, and what implications it might have on experimental design, data collection, and annotation? ii) What are the areas that might not have much benefit from taking a machine learning approach? This tutorial will present a number of acoustics/geophysics application examples along with their analogs in canonical machine learning applications, as an attempt to illustrate possibilities and limitations.