Testing

[rossviljoen]

The main approach to testing will (initially) involve checking the results of the sparse approximations when they should be equivalent to some known solution - for example:

• The SVGP posterior implemented here should be equivalent to the Titsias approximate posterior in AbstractGPs when the exact solution for the variational distribution is used (i.e. equations (11) & (12) from here).
• When the inducing inputs == the data inputs and with a Gaussian likelihood, the SVGP model should find the same solution as an exact GP after optimising the kernel hyperparameters.

[st--]

You can see this in GPflow in this notebook and this set of tests

[rossviljoen]

Started implementing this in #6

[rossviljoen]

These were implemented in #9