Unsupervised machine learning for acoustic tomography and event detection

We present acoustic tomography and event detection approaches based on unsupervised machine learning (ML). ML describes a set of techniques for automatically detecting and utilizing patterns in data. These techniques can roughly be categorized as either supervised or unsupervised learning. In supervised learning, an ML system (e.g., neural network (NN)) is trained to produce a desired output based on labeled training examples. In unsupervised learning, no labels are given and the task is to discover interesting or useful structure within the data. In acoustic and geophysical signal processing, often large training datasets are available, but have few labeled examples. We discuss how unsupervised ML can leverage training data without explicit labels to improve acoustic models. We first give an overview of the relevant ML theory. Next, we describe an approach to array tomography, called locally-sparse travel time tomography (LST). LST regularizes the inversion of travel time measurements for geophysical structure using dictionary learning (unsupervised). Dictionary learning is here used to obtain a dictionary of small-scale geophysical features that best represent the measurements. Finally, we discuss an approach to automated event detection in acoustic time series based on autoencoder NNs, which learn salient features from NN inputs with unsupervised learning.