Research map / Architectures and efficiency / Quantization
Unsupervised Neural Quantization for Compressed-Domain Similarity Search
Stanislav Morozov, Artem Babenko
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 |
|---|---|---|---|
| use UNQ instead of LSQ | LSQ · BigANN1M (8 bytes) | Recall@1: 29.2 -> 34.6 | Improved |
| use UNQ instead of LSQ | LSQ · Deep1M (16 bytes) | Recall@1: 41.1 -> 47.9 | Improved |
| add learned-decoder reranking to shallow LSQ | LSQ · BigANN1M (8 bytes) | Recall@1: 29.2 -> 30.3 | No clear effect |
| use UNQ instead of Catalyst + Lattice | Catalyst + Lattice · Deep1B (8 bytes) | Recall@1: 16.8 -> 14.5 | Hurt |
| use UNQ instead of Catalyst + Lattice | Catalyst + Lattice · BigANN1B (16 bytes) | Recall@1: 31.1 -> 38.3 | Improved |
| add decoder reranking of top candidates (ablation) | UNQ · BigANN1M (8 bytes) | Recall@1: 25.0 -> 34.6 | Improved |
| rerank exhaustively with d1 instead of top-500 candidates (ablation) | UNQ · BigANN1M (8 bytes) | Recall@1: 34.6 -> 34.6 | No clear effect |
| add triplet loss on the learned-space distance d2 (ablation) | UNQ · BigANN1M (8 bytes) | Recall@100: 95.7 -> 99.0 | Improved |
| add triplet loss on the learned-space distance d2 (ablation) | UNQ · BigANN1M (8 bytes) | Recall@1: 35.5 -> 34.6 | Hurt |
| use triplet loss only (no reconstruction term) instead of combined objective (ablation) | UNQ · BigANN1M (8 bytes) | Recall@1: 27.9 -> 34.6 | Hurt |
| add the CV^2 balance regularizer (ablation) | UNQ · BigANN1M (8 bytes) | Recall@1: 31.0 -> 34.6 | Improved |
| use hard (argmax straight-through) Gumbel-Softmax instead of soft (ablation) | UNQ · BigANN1M (8 bytes) | Recall@1: 33.8 -> 34.6 | Improved |
Closest papers by meaning
- Deep Visual-Semantic Quantization for Efficient Image Retrieval (CVPR 2017 · similarity 0.90)
- Self-Supervised Product Quantization for Deep Unsupervised Image Retrieval (ICCV 2021 · similarity 0.89)
- SUBIC: A Supervised, Structured Binary Code for Image Search (ICCV 2017 · similarity 0.89)
- Focused Quantization for Sparse CNNs (NeurIPS 2019 · similarity 0.88)
- Network Memory Footprint Compression Through Jointly Learnable Codebooks and Mappings (ICLR 2024 · similarity 0.88)
- Qinco2: Vector Compression and Search with Improved Implicit Neural Codebooks (ICLR 2025 · similarity 0.86)
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.
