Seminars & Colloquia Calendar

Download as iCal file

DIMACS Theory of Computing Seminar

Practical Data-Dependent Metric Compression with Provable Guarantees

Ilya Razenshteyn, Columbia University

Location:  CoRe 301
Date & time: Wednesday, 11 October 2017 at 11:00AM - 12:00PM

Abstract:  How well can one compress a dataset of points from a high-dimensional space while preserving pairwise distances? Indyk and Wagner have recently obtained almost optimal bounds for this problem, but their construction (based on hierarchical clustering) is not practical. In this talk, I will show a new practical, quadtree-based compression scheme, whose provable performance essentially matches that of the result of Indyk and Wagner.

 In additional to the theoretical results, we will see experimental comparison of the new scheme and Product Quantization (PQ)--one of the most popular heuristics for distance-preserving compression--on several datasets. Unlike PQ and other heuristics that rely on the clusterability of the dataset, the new algorithm ends up being more robust.

 The talk is based on a joint work with Piotr Indyk and Tal Wagner.

Special Note to All Travelers

Directions: map and driving directions. If you need information on public transportation, you may want to check the New Jersey Transit page.

Unfortunately, cancellations do occur from time to time. Feel free to call our department: 848-445-6969 before embarking on your journey. Thank you.