Sample Amplification: Increasing Dataset Size Even When Learning is Impossible

Explore the concept of "sample amplification" in this 59-minute USC Probability and Statistics Seminar talk. Delve into the intriguing question of whether generating new samples from a distribution always requires learning it completely. Discover how it's possible to output a larger set of datapoints that are indistinguishable from independent draws from an unknown distribution, even when the initial sample size is too small for accurate learning. Examine upper and lower bounds on sample amplification for various distribution families, including discrete distributions, Gaussians, and continuous exponential families. Learn about the joint research findings of Vatsal Sharan, Brian Axelrod, Yanjun Han, Shivam Garg, and Greg Valiant in this thought-provoking presentation on expanding dataset sizes in challenging learning scenarios.