Daniel Dodd

Post #370, refactor the BasisFunctionComputation (that sadly is dense) such that it uses a low rank linear operator - so O(MN^2) computational cost. Ensure this is the default computation for...

Switch between them like: ```python nat_q = NaturalVarationalGaussian(...) exp_q = nat_q.to_exp() # ExpectationVarationalGaussian exp_q.to_nat() # Natural Variational Gaussian ```

# Feature Request Post Python 3.10 bump #370, to avoid issues with default values, we should use the KW_ONLY, e.g., for `names` most notably this would be great for default...

Following the integration of CoLA (#370). It would be great to add a Konecker Kernel + computation. ```python from gpjax.kernels import CombinationKernel, RBF # Code inherits from CombinationKernel or ProductKernel...

For the exponential link (which we have) the poisson likelihood variational expectation / expected log likelihood can be computed in closed form. Proposal is to add a `ExponentialPoissonIntegrator` to compute...

Add Monte Carlo integrator to compute the variational expectation / expected log likelihood: $$\mathbb{E}_{q(\mathbf{f})} \log p(y|\mathbf{f}) \approx \frac{1}{N} \sum_i \log p(y|\mathbf{f}^{(i)}), \qquad \mathbf{f}^{(i)} \sim q(\mathbf{f}). $$ ```python class MonteCarloIntegrator(AbstractIntegrator): ......

dev: VI: Refactor natural gradients, add dual parameterisation, add pseudo data.

1

Though mentioned in meeting minutes, making my development on this public. Will tick things off as and when they are completed. ## Checklist Add: - [ ] Dual parameterisation of...

Improve / add unit tests for `GaussianDistribution`. (1) We need to test shape checking done by the function `check_shapes`. (2) Remove `distrax ` checking for equality with e.g., likelihood value...

Use Choleksy decomposition instead of `jnp.linalg.inv` in Multivariate Gaussian Distributions.

1

Thanks for creating such a nice library. I have experienced numerical instability in the Multivariate Gaussian Distributions. For better numerical stability, I wondered if it would be better replace e.g.,...