Hey guys,
I’ve been able to successfully implement my own sparse latent gaussian process model with an arbitrary number of latent dimensions (with inducing inputs). However, I want to allow correlation over time of the latent state. This paper gives an idea on how to do it:
(Variational Gaussian Process Dynamical Systems - see section 2.2)
https://arxiv.org/pdf/1107.4985
However, the parametrization of the guide appears to require the inversion of an NxN matrix, which isn’t feasible for any reasonable sized problem (which is obviously why the inducing-input approach was developed).
Am I missing something here? How can this be done?
Mike