3 ms·
> This connection is interesting because it implies that differential privacy techniques applied to machine learning must in turn result in more generalizable m
by mehrdadn 9y ago
> This connection is interesting because it implies that differential privacy techniques applied to machine learning must in turn result in more generalizable models, which is very useful in general.
I'm confused. Wouldn't a model that produces completely random results be differentially private but not generalizable? Or is giving more random results considered generalization?