A Bi-metric Framework for Fast Similarity Search

Explore a 38-minute lecture on a novel bi-metric framework for designing efficient nearest neighbor data structures. Delve into Piotr Indyk's presentation from the Massachusetts Institute of Technology, delivered at the Simons Institute as part of the Extroverted Sublinear Algorithms series. Discover how this framework utilizes two dissimilarity functions: an accurate but computationally expensive ground-truth metric and a less precise but faster proxy metric. Learn about the theoretical and practical applications of this approach, including its implementation in popular algorithms like DiskANN and Cover Tree. Examine how the framework achieves high accuracy while limiting calls to both metrics, and understand its potential for improving text retrieval tasks. Gain insights into the accuracy-efficiency tradeoffs observed in the MTEB benchmark and the advantages of this method over alternatives such as re-ranking.

A Bi-metric Framework for Fast Similarity Search