AugmentedGPLikelihoods.jl

AugmentedGPLikelihoods.jl copied to clipboard

This pull request changes the compat entry for the `MeasureTheory` package from `0.16 - 0.18` to `0.16 - 0.18, 0.19`. This keeps the compat entries for earlier versions. Note: I...

github-actions[bot]

Refresh reference to double bound when available

2

Once the thesis is out or a paper is published put in the Heteroscedastic Gaussian docs

theogf

There seems to be an issue with the optimal variational distribution for q(Omega). It comes either from the computation of the ELBO or from the optimal variational distribution itself.

theogf

- [x] Student-T #10 - [ ] BayesianSVM - [x] LaplaceLikelihood #15 - [x] Categorical{

theogf

In some cases the prior and tilt cannot be splitted (in Categorical for example), we should still have a general approach for it.

theogf

How to do hyperparameter tuning

7

I tried to AD `aug_elbo` in the `NegBinomialLikelihood` example, i.e. (removed unnecessary bits), purposefully avoiding ParameterHandling.jl and trying only with `ForwardDiff.gradient` ```julia # # Negative Binomial # We load all...

simsurace

- [ ] Heteroscedastic Laplace Likelihood

theogf


theogf

theogf

There can be differentiation issues (with Zygote) regarding the auxiliary variables, especially with `expected_logtilt`. #20 Solved this problem for now with a relatively generic solutions but more tests are needed.

theogf