Department of Computer Science | Institute of Theoretical Computer Science | CADMO

Theory of Combinatorial Algorithms

Prof. Emo Welzl and Prof. Bernd Gärtner

Mittagsseminar (in cooperation with M. Ghaffari, A. Steger and B. Sudakov)

Mittagsseminar Talk Information

Date and Time: Thursday, March 15, 2018, 12:15 pm

Duration: 30 minutes

Location: CAB G51

Speaker: David Steurer

Improved clustering and robust moment estimation via sum-of-squares

We develop efficient algorithms for clustering and robust moment estimation. Via convex sum-of-squares relaxations, our algorithms can exploit bounds on higher-order moments of the underlying distributions. In this way, our algorithms achieve significantly stronger guarantees than previous approaches.

For example, our clustering algorithm can learn mixtures of k spherical Gaussians in time nO(t) as long as the means have minimum separation t1/2k1/t for any t ≥ 4, significantly improving over the minimum separation k1/4 required by the previous best algorithm due to Vempala and Wang (JCSS’04).

Given corrupted samples that contain a constant fraction of adversarial outliers, our moment estimation algorithms can approximate low-degree moments of unknown distributions, achieving guarantees that match information-theoretic lower-bounds for a broad class of distributions. These algorithms allow us to enhance in a black-box way many previous learning algorithms based on the method of moments, e.g., for independent component analysis and latent dirichlet allocation, such that they become resilient to adversarial outliers.

Based on joint works with Pravesh Kothari and Jacob Steinhardt.


Upcoming talks     |     All previous talks     |     Talks by speaker     |     Upcoming talks in iCal format (beta version!)

Previous talks by year:   2018  2017  2016  2015  2014  2013  2012  2011  2010  2009  2008  2007  2006  2005  2004  2003  2002  2001  2000  1999  1998  1997  1996  

Information for students and suggested topics for student talks


Automatic MiSe System Software Version 1.4803M   |   admin login