Currently, Mxnet has no differentiable operators for the log probability density of standard distributions. For discrete distributions, this is trivial. For a vector a single variate Gaussian distributions, implementing such an operator is fairly straightforward. However, implementing an operator for a multivariate Gaussian log density is not so straightforward.

In particular, unless I’m missing something, to implement one we would need access to more linear algebra operations for determinants and inverses. While Mxnet seems to have some matrix operations, I think they’re actually too narrow in scope to support what we would need for a multivariate Gaussian log density, but my linear algebra could use some brushing up, so it’s possible I’m just not seeing how to use what is there to do it.

Has anyone here implemented one themselves, or see how it could be done?