Research map / Architectures and efficiency / RNNs and normalization
Revisiting the Inverted Indices for Billion-Scale Approximate Nearest Neighbors
Dmitry Baranchuk, Artem Babenko, Yury Malkov
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 |
|---|---|---|---|
| add grouping (L=64 subcentroids) to the inverted index (ablation) | Inverted index K=2^20 · DEEP1B | average distance to closest centroid: 0.282 -> 0.255 | Improved |
| add grouping to the inverted multi-index (ablation) | IMI K=2^10 · DEEP1B | average distance to closest centroid: 0.415 -> 0.385 | Improved |
| increase the inverted index codebook from K=2^18 to 2^22 | Inverted index · DEEP1B | average distance to closest centroid: 0.315 -> 0.259 | Improved |
| increase the IMI codebook from K=2^13 to 2^15 | Inverted multi-index · DEEP1B | average distance to closest centroid: 0.345 -> 0.305 | Improved |
| increase subcentroids per region from L=32 to 64 (ablation) | IVFOADC+Grouping+Pruning · DEEP1B (16-byte codes) | Recall@1: 0.417 -> 0.433 | Improved |
| increase subcentroids per region from L=64 to 128 (ablation) | IVFOADC+Grouping+Pruning · DEEP1B (16-byte codes) | Recall@1: 0.433 -> 0.441 | Improved |
| increase subcentroids per region from L=64 to 128 (runtime) (ablation) | IVFOADC+Grouping+Pruning · DEEP1B (16-byte codes) | search time (ms): 1.28 -> 1.48 | Hurt |
| use inverted index with large codebook, grouping and pruning instead of inverted multi-index | IVFOADC+G+P vs O-Multi-D-OADC · DEEP1B (16-byte codes) | Recall@1: 0.397 -> 0.452 | Improved |
| use inverted index with grouping and pruning instead of inverted multi-index | IVFOADC+G+P vs O-Multi-D-OADC · SIFT1B (16-byte codes) | Recall@1: 0.360 -> 0.405 | Improved |
| use inverted index with grouping and pruning instead of Multi-LOPQ | IVFOADC+G+P vs Multi-LOPQ · SIFT1B (16-byte codes) | Recall@1: 0.454 -> 0.405 | Hurt |
Closest papers by meaning
- Efficient Indexing of Billion-Scale Datasets of Deep Descriptors (CVPR 2016 · similarity 0.89)
- Product Quantizer Aware Inverted Index for Scalable Nearest Neighbor Search (ICCV 2021 · similarity 0.88)
- Adaptive Retrieval and Scalable Indexing for k-NN Search with Cross-Encoders (ICLR 2024 · similarity 0.85)
- Semi-Parametric Retrieval via Binary Bag-of-Tokens Index (ICLR 2025 · similarity 0.84)
- Discovering Data Structures: Nearest Neighbor Search and Beyond (NeurIPS 2025 · similarity 0.84)
- Graph-Based Algorithms for Diverse Similarity Search (ICML 2025 · similarity 0.84)
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.
