The kernel based methods clone the training data

[AtheMathmo]

trafficstars

This is very space inefficient. Instead of doing this we could provide lifetimes and use references to the training data.

Perhaps this should be tackled as part of a wider attempt at space optimization as we introduce views etc. into the matrices.

[k0pernicus]

It seems kernel based methods now use references to access and manipulate data. Is this issue resolved, or could you provide an example please?