Semi-Supervised Learning for Seismic Monitoring Applications

Date: 4/26/2019

Time: 10:45 AM

Room: Elliott Bay

The impressive performance that Deep Neural Networks demonstrate on a range of seismic monitoring tasks is largely due to the availability of event catalogs that have been manually curated over many years or decades. However, the quality, duration, and availability of seismic event catalogs varies significantly across the range of monitoring operations, regions, and objectives. Semi-Supervised Learning (SSL) methods provide a framework to leverage the abundance of unreviewed seismic data, that would otherwise go unused for learning a variety of target tasks. We apply two recently published SSL algorithms (Mean-Teacher and Virtual Adversarial Training) on seismic event classification to examine how unlabeled data can enhance model performance. We explore how SSL techniques can close the generalization gap when the input and target distributions differ as well as when they are approximately equivalent, the intended use case of SSL. We provide discussion for when these techniques work well and cautions for when using SSL for monitoring in regions where no ground truth is available.

Presenting Author: Lisa Linville

Authors