BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//wp-events-plugin.com//7.2.3.1//EN
TZID:America/New_York
X-WR-TIMEZONE:America/New_York
BEGIN:VEVENT
UID:521@fds.yale.edu
DTSTART;TZID=America/New_York:20240408T160000
DTEND;TZID=America/New_York:20240408T170000
DTSTAMP:20250916T142132Z
URL:https://fds.yale.edu/events/sds-seminar-santosh-vempala-georgia-tech/
SUMMARY:S&DS Seminar: Santosh Vempala (Georgia Tech)\, "Robust Statistics i
 n High Dimension"
DESCRIPTION:3:30pm - Pre-talk meet and greet teatime - 219 Prospect Street\
 , 13 floor\, there will be light snacks and beverages in the kitchen area.
 Information and Abstract: The goal of robust statistics is to find accurat
 e estimates of statistical parameters despite adversarial corruptions of d
 ata (an eps fraction of data is arbitrarily corrupted by an adversary). Fo
 r example\, the median is a robust estimator of the mean for a Gaussian di
 stribution. While this topic has been studied since at least 1960 (Huber\;
  Tukey)\, proposed solutions either had error that scaled polynomially wit
 h the dimension or had running times that scaled exponentially.  In 2016\,
  concurrent papers [DKKLMS\; LRV] gave efficient\, robust algorithms for m
 ean and covariance estimation of high-dimensional Gaussians (and generaliz
 ations). Since then\, there has been steady progress on robust estimation 
 and learning algorithms. In this talk\, we will discuss two results for cl
 assical statistics problems: (1) a polynomial-time algorithm for robustly 
 learning a mixture of k arbitrary Gaussians (with Bakshi\, Diakonikolas\, 
 Jia\, Kane and Kothari\, 2022)\, which relies on a robust partial clusteri
 ng algorithm and robust tensor decomposition\, both of independent interes
 t as algorithmic tools and (2) a polynomial-time algorithm for learning an
  affine transformation of a unit hypercube\, a basic setting of Independen
 t Component Analysis (with Jia and Kothari\, 2023). The latter is a proble
 m that information-theoretically cannot be solved from robust estimates of
  moments (unlike essentially all known solvable robust estimation problems
 )\, and our algorithm provides new certificates for affine transformations
 \, immune to adversarial noise\; the main idea is a robust version of grad
 ient descent.\n\nClick for more information\n
CATEGORIES:FDS Events,Statistics & Data Science Seminar,Seminar Series
END:VEVENT
BEGIN:VTIMEZONE
TZID:America/New_York
X-LIC-LOCATION:America/New_York
BEGIN:DAYLIGHT
DTSTART:20240310T030000
TZOFFSETFROM:-0500
TZOFFSETTO:-0400
TZNAME:EDT
END:DAYLIGHT
END:VTIMEZONE
END:VCALENDAR