Implementation and derivation of "Variational Bayesian inference for a nonlinear forward model." [Chappell et al. 2008] for arbitrary, user-defined model errors.
The current implementation always infers the noise. The option to not do that and instead provide a determinstic noise-sigma (as common in likelihood functions) is missing.
IMO this could be implemented by just skipping the update equations 21-22 (or 30-31 for multiple noises). Then, in the remaining update equations, the noise is only used as "scale x shape" which corresponds to the mean of this distribution, which could be provided instead.
Discussion:
Are my thoughts correct?
The free energy terms include the noise shape and noise scale separately. How to handle that?
The current implementation always infers the noise. The option to not do that and instead provide a determinstic noise-sigma (as common in likelihood functions) is missing.
IMO this could be implemented by just skipping the update equations 21-22 (or 30-31 for multiple noises). Then, in the remaining update equations, the noise is only used as "scale x shape" which corresponds to the mean of this distribution, which could be provided instead.
Discussion: