Fast Distance for Approximate NearestNeighbor Search: A Unified Survey

The Hong Kong University of Science and Technology (Guangzhou)

数据科学与分析学域

PhD Qualifying Examination

By Mr. JING, Liuchang

摘要

Approximate nearest neighbor search (ANNS) is a core primitive in large-scale retrieval,

recommendation, and RAG. Among various ANNS methods, the distance computation,

which evaluates the similarity between a query and every candidate vector, is prohibitively

expensive and shown to be the critical performance bottleneck. Two complementary

families of techniques are proposed to accelerate this step: Fast Distance Calculation

(FDCal) methods, which compress database vectors via vector quantization to enable

cheap approximate distance lookups; and Fast Distance Comparison (FDCom) methods,

which compute bounds on approximation distances to prune unpromising candidates

before any expensive computation. In this survey, we introduce a unified interval

estimate framework in which both families produce bounds [δ ˆ−, δ ˆ+] on true distance

with reliability ρ and cost C. The (bound, reliability, cost) triple (ε, ρ, C) unifies their

design tradeoffs and reveals FDCal as a constrained special case of FDCom that explicitly

minimizes the approximation error ε. Then we present taxonomies for both families and

their integration with ANNS indexes.

PQE Committee

• Chair: Prof. CHU, Xiaowen
• Prime Supervisor: Prof. WANG, Wei
• Co-Supervisor: Prof. LI, Lei
• Examiner: Prof. ZHANG, Yongqi