I am trying to train an autoencoding neural network (autoencoder) to reconstruct seismograms. Previous studies employing this technique (e.g. Valentine & Trampert, 2012) used an L2 (mean squared error) loss function to assess the network's performance, but I'm wondering if there are better loss functions out there for this type of data. Seismograms are rather spurious, in the sense that the signal rapidly oscillates around zero, which makes it hard to get a good point-by-point comparison. For instance, if the reconstructed seismogram has a slight time lag but is otherwise a perfect reconstruction of the original input (see image), the L2 loss could be enormous due to this slight misalignment.

seismogram reconstruction

My question: how would you evaluate the quality of the reconstruction in terms of the approximate position, duration, amplitude, etc.?

  • $\begingroup$ I don't think this is a computer science question; the loss function has to be based on the needs of the application (e.g., on the physics of earthquakes), and that's not a matter of computer science but a matter of physics or geology or whatever your application domain on. $\endgroup$ – D.W. Mar 27 '19 at 18:32
  • $\begingroup$ Well, the application would be (lossy) signal compression by means of an autoencoder. This isn't really a geophysics question, the input signal could (theoretically) be any oscillatory time series. $\endgroup$ – MPA Mar 27 '19 at 21:11
  • $\begingroup$ I'm saying that the loss function / metric for evaluation has to be informed by what you plan to use this auto encoder for -- it's not a matter of computer science. Given a loss function, how to train an autoencoder might be a CS question, but choosing the metric for evaluation isn't. There's no one right answer. $\endgroup$ – D.W. Mar 27 '19 at 21:18

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.