This paper introduces variants of beam search for DiskANN, a state-of-the-art method for large-scale approximate nearest-neighbor search. The research unifies and generalizes existing theoretical results on graph-based search, providing improved approximation bounds. Practical experiments show that a novel beam predictor, trained using fine-grained analysis, significantly reduces query latency on SSD-resident indices while maintaining high recall. AI
IMPACT Improves efficiency for large-scale data retrieval, potentially impacting AI systems reliant on fast similarity searches.
RANK_REASON This is a research paper detailing theoretical advancements and experimental results for an approximate nearest-neighbor search algorithm. [lever_c_demoted from research: ic=1 ai=0.7]
Read on arXiv cs.IR (Information Retrieval) →
- Al-Jazzazi
- CBS
- DiskANN
- Gollapudi
- Hierarchical Navigable Small-World graphs
- Indyk
- Nuclear Suppliers Group
- Xu
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →