Non-Negative Tensor Factorization for Interpretable Unsupervised Signal Discovery in Continuous Seismic Data
Date: 4/24/2019
Time: 02:15 PM
Room: Elliott Bay
Discovering novel signals in continuous seismic data, particularly those that aren't characterized by common attributes like transience, high amplitudes, or narrow frequencies, is difficult. Unsupervised machine learning offers ways to discover novel signals, but many techniques suffer from the problem of non-interpretability. Non-negative tensor factorization (NTF) is an unsupervised learning approach that can extract hidden features from within data, subject to non-negativity constraints, which makes these features more physically interpretable than comparable techniques like principal component analysis. We present results of applying NTF to continuous seismic polarization and power spectrum data, towards physically interpretable automated feature discovery.
Presenting Author: Benjamin T. Nebgen
Authors
Benjamin T Nebgen bnebgen@lanl.gov Los Alamos National Laboratory, Los Alamos, New Mexico, United States Presenting Author
|
Jonathan K MacCarthy jkmacc@lanl.gov Los Alamos National Laboratory, Los Alamos, New Mexico, United States Corresponding Author
|
Boian Alexandrov boian@lanl.gov Los Alamos National Laboratory, Los Alamos, New Mexico, United States |
Non-Negative Tensor Factorization for Interpretable Unsupervised Signal Discovery in Continuous Seismic Data
Category
Machine Learning in Seismology