Research map / Architectures and efficiency / Quantization
Approximate Search with Quantized Sparse Representations
Himalaya Jain, Patrick Pérez, Rémi Gribonval, Joaquin Zepeda, Hervé Jégou
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 quantized scalar coefficients to residual quantization (Qα-RVQ vs RVQ) | RVQ · SIFT1M 64 bits (un-normalized) | squared reconstruction error: 22170 -> 22053 | Improved |
| add quantized scalar coefficients to residual quantization (Qα-RVQ vs RVQ) | RVQ · GIST1M 64 bits (un-normalized) | squared reconstruction error: 0.6986 -> 0.6537 | Improved |
| add quantized coefficients to PQ (Qα-PQ vs PQ) at equal code size | PQ · SIFT1M 64 bits (un-normalized) | squared reconstruction error: 23515 -> 25859 | Hurt |
| add quantized coefficients to PQ (Qα-PQ vs PQ) at equal code size | PQ · GIST1M 64 bits (un-normalized) | squared reconstruction error: 0.7121 -> 0.7224 | Hurt |
| use OPQ instead of PQ inside Qα-PQ (Qα-OPQ) (ablation) | Qα-PQ · SIFT1M 64 bits (Euclidean aNN) | recall R@1: 0.204 -> 0.227 | Improved |
| use OPQ instead of PQ inside Qα-PQ (Qα-OPQ) (ablation) | Qα-PQ · SIFT1M 64 bits (Euclidean aNN) | recall R@100: 0.900 -> 0.927 | Improved |
| apply OPQ to plain PQ (ablation) | PQ · SIFT1M 64 bits (Euclidean aNN) | recall R@1: 0.228 -> 0.243 | Improved |
| Qα-RVQ instead of RVQ with IMI T=100K | RVQ · BIGANN 1B IMI T=100K | recall R@1: 0.181 -> 0.227 | Improved |
| Qα-PQ instead of PQ-64 with IMI T=100K | PQ-64 · BIGANN 1B IMI T=100K | recall R@1: 0.170 -> 0.200 | Improved |
| Qα-PQ instead of PQ for encode time (half the codewords, inner products) | PQ · SIFT1M 64 bits relative encoding time | relative encode time: 1 -> 0.206 | Improved |
| Qα-PQ instead of PQ for search time | PQ · SIFT1M 64 bits relative search time | relative search time: 1 -> 1.867 | Hurt |
| add candidate pruning W'c=128 to Qα-RVQ in IVF (ablation) | Qα-RVQ · BIGANN 1B IVF | search time and recall: 10% faster than PQ with no performance loss | Improved |
Closest papers by meaning
- Composite Quantization (TPAMI 2018 · similarity 0.89)
- Compressive Quantization for Fast Object Instance Search in Videos (ICCV 2017 · similarity 0.88)
- Minimizing FLOPs to Learn Efficient Sparse Representations (ICLR 2020 · similarity 0.88)
- Polysemous Codes (ECCV 2016 · similarity 0.88)
- Qinco2: Vector Compression and Search with Improved Implicit Neural Codebooks (ICLR 2025 · similarity 0.87)
- Beyond Matryoshka: Revisiting Sparse Coding for Adaptive Representation (ICML 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.
