#memorization

NVIDIA pointed to an ICML paper that separates unintended memorization from real generalization. Its headline estimate, about 3.6 bits per parameter, gives privacy and scaling debates a concrete capacity number.