Research map / Architectures and efficiency / RNNs and normalization
SPANN: Highly-efficient Billion-scale Approximate Nearest Neighborhood Search
Qi Chen, Bing Zhao, Haidong Wang, Mingqin Li, Chuanjie Liu, Zengzhong Li, Mao Yang, Jingdong Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Hierarchical balanced clustering for centroids instead of random or hierarchical KMeans (ablation) | SPANN · SIFT1M recall@1 and recall@10 | recall vs latency | Improved |
| Increase the number of centroids toward 16% of points (ablation) | SPANN · SIFT1M | recall vs latency | Improved |
| Increase centroids beyond 16% of points (ablation) | SPANN · SIFT1M | recall vs latency | No clear effect |
| Use more than one closure replica per vector (ablation) | SPANN · SIFT1M | recall vs latency | Improved |
| Increase closure replicas above 8 (ablation) | SPANN · SIFT1M | recall vs latency | No clear effect |
| Add query-aware dynamic pruning of posting lists (ablation) | SPANN · SIFT1M | query latency at same recall | Improved |
| Replace random partition with multi-constraint balanced clustering (ablation) | Distributed SPANN, 32 partitions · SPACEV1B | machines dispatched per query: 32 -> 9 | Improved |
| Add closure assignment to balanced clustering (ablation) | Distributed SPANN, 32 partitions · SPACEV1B | machines dispatched per query: 9 -> 8 | Improved |
| Add query-aware dynamic pruning online (ablation) | Distributed SPANN, 32 partitions · SPACEV1B | machines dispatched per query: 8 -> 6.3 | Improved |
| Closure assignment increases vectors stored per machine | Distributed SPANN · SPACEV1B, 8/16/32 partitions | vectors per machine: 20% | Hurt |
| Use more machines (8, 16, 32 partitions) with balanced partitioning vs random partition | Distributed SPANN · SPACEV1B query accesses per machine | query accesses per machine | Improved |
| SPANN instead of in-memory graph ANNS (NSG, HNSW, SCANN, NGT, N2) | SIFT1M · SIFT1M recall@1, recall@10 | VQ capacity | Improved |
Closest papers by meaning
- DiskANN: Fast Accurate Billion-point Nearest Neighbor Search on a Single Node (NeurIPS 2019 · similarity 0.93)
- LotusFilter: Fast Diverse Nearest Neighbor Search via a Learned Cutoff Table (CVPR 2025 · similarity 0.87)
- Optimistic Query Routing in Clustering-based Approximate Maximum Inner Product Search (NeurIPS 2025 · similarity 0.86)
- Results of the Big ANN: NeurIPS’23 competition (NeurIPS 2025 · similarity 0.86)
- Graph-Based Algorithms for Diverse Similarity Search (ICML 2025 · similarity 0.85)
- Sort Before You Prune: Improved Worst-Case Guarantees of the DiskANN Family of Graphs (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
