The spectral mixture kernel can be difficult to train with multiple components, because there is a degeneracy between the different peaks, which is exacerbated by having multiple wavelengths. We can make this easier for some cases by just using a SMK along the time axis and using an RBF or Matern kernel along the wavelength axis instead for simpler inference.
The spectral mixture kernel can be difficult to train with multiple components, because there is a degeneracy between the different peaks, which is exacerbated by having multiple wavelengths. We can make this easier for some cases by just using a SMK along the time axis and using an RBF or Matern kernel along the wavelength axis instead for simpler inference.