Is there some effective way to specifiy that there is no noise present in the data for the marginal_likelihood method? I tried to set the parameter ‘noise = 0’, but pm.sample does not seem to be particularly happy about that.
Any tips would be greatly appreciated!
In Gaussian Process the
marginal_likelihood also stabilized the covariance function, so if you just set it to 0 sometimes you have numerical problems.
A quick workaround is to set noise to a small value, but a more proper treatment is to revisit the priors for covariance functions to use more informative prior that constructs more stable covariance matrix.
One thing to keep in mind too is that
gp.Marginal assumes that there is Gaussian noise in the observations. The difference between
Latent is that for
f is integrated out (since multivariate normals are conjugate). So if there is no noise, you could consider using Latent. Assuming the noise is very very small though is effectively the same thing and will probably behave better numerically.