Skip to main content
Browse by:
GROUP

Stein's Paradox and Multi-task Averaging (to appear, JMLR 2014)

Event Image
Thursday, September 11, 2014
11:45 am - 12:45 pm
Maya Gupta-Google Research
Data Seminar

Stein showed that you could do better estimating the means of different, independent random variables if you mixed their samples together, that is, you can do a better job (in total mean-squared error sense) of estimating the price of tea in China and batting averages if you mix both data-sets appropriately. We re-frame this problem as taking a multi-task approach to estimating multiple means at once, setting up the problem in the standard machine learning structural risk minimization framework. This formulation is intuitive as a regularization, and has some nice theoretical properties. We'll show improved performance over James-Stein estimation for estimating students' grades, Amazon customer reviews, and the length of kings' reigns. Bio: Gupta joined Google Research in 2012. Before Google, Gupta was an Assoc Prof of Electrical Engineering at U Washington. Her research group focused on statistical learning algorithms for signal processing and color image processing. In 2007, Gupta received the PECASE award from Pres. George Bush for her work in classifying uncertain (e.g. random) signals, and received the 2007 Office of Naval Research YIP Award. Her Ph.D. in Electrical Engineering is from Stanford University (2003), where she was a National Science Foundation Graduate Fellow and worked with Bob Gray, Rob Tibshirani, and Richard Olshen. Before 2003, Gupta worked for Ricoh Research, NATO's Undersea Research Center, HP R&D, AT&T Labs, and Microsoft.

Type: LECTURE/TALK
Contact: Kathy Peterson