Using a lazy representation of the kernel matrix

[sethaxen]

Currently this package eagerly constructs and uses a dense kernel matrix. However, iterative GP methods are very useful when the data is large enough that the kernel matrix can't even fit into memory. It would be better to construct and use a lazy representation of the kernel matrix.

I've started work on adding this functionality to the KernelFunctions API: JuliaGaussianProcesses/KernelFunctions.jl#515

[SebastianCallh]

Very good comment, and I completely agree. Because of #5 I have been hesitant to introduce additional moving parts since it would probably make it more difficult to solve.

Looks like this should be fairly straightforward after JuliaGaussianProcesses/KernelFunctions.jl#515 lands? So perhaps it makes sense to wait for the linked PRs to be resolved before tackling this?

[sethaxen]

Looks like this should be fairly straightforward after JuliaGaussianProcesses/KernelFunctions.jl#515 lands? So perhaps it makes sense to wait for the linked PRs to be resolved before tackling this?

Agreed! This issue is mostly for tracking the feature.