Research talk: Approximate nearest neighbor search systems at scale

Speaker: Harsha Simhadri, Principal Researcher, Microsoft Research India Building deep learningbased search and recommendation systems at internet scale requires a complete redesign of the search index. Key to this redesign is a fast, accurate, and costefficient indexing system for approximate nearest neighbor search. In this talk, well present our recent advances in this space, including the DiskANN and FreshDiskANN systems and the underlying algorithms. These algorithms present an orderofmagnitude improvement in scale and costofoperation over the state of the art and are a first of their kind at effectively using solidstate drives (SSDs) to serve at interactive (milliseconds) latencies. In addition, they provide faster inmemory search than other graph indices, like HNSW, and support realtime concurrent insertions and deletions to SSDresident indices without losing recall. Well provide an overview their applicability to various product scenarios and highlight directions for further researc