CS Talk - Manolis Zampetakis, MIT
Host: Yang Cai
Title: Computationally and Statistically Efficient Truncated Statistics
Abstract:
Censoring and truncation occur when data falling outside of a subset of the population are not observable. In practice, it often arises as a result of saturation of measurement devices, experimental design, and legal or privacy constraints preventing the use of some of the data. Such phenomena have been known to affect experimental results in a counterintuitive way, as per Berkson’s paradox.
In our recent work, we provide the first provably computationally and statistically efficient methods accomplishing the fundamental task of statistical estimation for the entire population out of exclusively censored data. Our first result [w/ Daskalakis, Gouleakis, Tzamos FOCS’18] assumes that the population follows a multi-dimensional normal distribution and the survival set is known. In follow-up works, we have extended our result to the case of censored linear [w/ Daskalakis, Gouleakis, Tzamos COLT’19], logistic and probit regression [w/ Daskalakis, Ilyas, Rao ‘19] and we have also explored the case of unknown survival set [w/ Kontonis, Tzamos FOCS’19].
Bio:
Manolis is a Ph.D. Student in the Theory of Computation Group at MIT working on theoretical problems in: machine learning and learning theory, complexity theory, and algorithmic game theory. Before MIT, Manolis was an undergraduate student at National Technical University of Athens. He has been an intern at Google Research NYC, Yahoo! Research NYC and Microsoft Research New England. His graduate studies have been supported by a Google Ph.D. Fellowship.