Identifying wrong labels

[janfreyberg]

Giles mentioned that it would be useful to have some way of identifying "wrong" labels – points that might be mislabelled – based on leave-one-out training and testing model accuracy.

It would be nice to provide an interface for that!

[janfreyberg]

Just some more thoughts:

• Influence functions could work well for this
• Distance metrics in some dimensionality-reduced space might help with not having to re-train models all the time