HR: 1340h
AN: NG43B-0571    [Abstracts]
TI: Support Vector Machines for Geophysical Inversion
AU: * Kuzma, H A
EM: hkuzma@berkeley.edu
AF: University of California, Berkeley, P.O. Box 987, Truckee, CA 96160
AU: Rector, J W
EM: jwrector@lbl.gov
AF: University of California, Berkeley, P.O. Box 987, Truckee, CA 96160
AB: Support Vector Machines (SVMs) trained using physically realistic models can be used to approximate the results of geophysical inversion. SVMs for inversion compare field data to synthetic data and interpolate between models. If the inverse problem is non-unique, the results of an SVM depend heavily on the training data. An analogy can be made between regularization, Bayesian priors and choice of training data. Training an SVM requires solving an easy problem that scales with the size of the training set. Geophysical SVMs give good performance with relatively small training sets. Because SVMs are fast, confidence intervals can be computed using Jackknife or Bootstrap methods. Examples are presented from seismic inversion. An easy-to-use MATLAB interface has been developed for general inverse problems
DE: 3260 Inverse theory
DE: 3275 Uncertainty quantification (1873)
DE: 4450 Nonlinear maps
DE: 9810 New fields (not classifiable under other headings)
DE: 9820 Techniques applicable in three or more fields
SC: Nonlinear Geophysics [NG]
MN: Fall Meeting 2005