High-dimensional separability for one- and few-shot learning