Fall 2018

Importance Sampling in High Dimensions via Hashing

Wednesday, November 28th, 2018 9:30 am10:30 am

Add to Calendar


Moses Charikar (Stanford University)

Locality sensitive hashing (LSH) is a popular technique for nearest neighbor search in high dimensional data sets. Recently, a new view at LSH as a biased sampling technique has been fruitful for density estimation problems in high dimensions. Given a set of points and a query point, the goal (roughly) is to estimate the density of the data set around the query. One way to formalize this is by kernel density estimation: Given a function that decays with distance and represents the "influence" of a data point at the query, sum up this influence function over the data set. Yet another way to formalize this problem is by counting the number of data points within a certain radius of the query. While these problems can easily be solved by making a linear pass over the data, this can be prohibitive for large data sets and multiple queries. Can we preprocess the data so as to answer queries efficiently? This talk will survey several recent papers that use locality sensitive hashing to design unbiased estimators for such density estimation problems and their extensions. This talk will survey joint works with Arturs Backurs, Piotr Indyk, Vishnu Natchu, Paris Syminelakis and Xian (Carrie) Wu.