Research map / Learning theory and algorithms / Prediction and decision losses
Polysemous Codes
Matthijs Douze, Hervé Jégou, Florent Perronnin
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 |
|---|---|---|---|
| optimize index assignment (distance loss) before binary comparison (ablation) | PQ/binary · SIFT1M | recall@1: 0.036 -> 0.107 | Improved |
| optimize index assignment (distance loss) before binary comparison (ablation) | PQ/binary · SIFT1M | recall@100: 0.129 -> 0.503 | Improved |
| optimize index assignment (distance loss) for Hamming filtering (ablation) | PQ/dual · SIFT1M | recall@1: 0.312 -> 0.441 | Improved |
| optimize index assignment (distance loss) for Hamming filtering (ablation) | PQ/dual · FYCNN1M | recall@100: 0.522 -> 0.804 | Improved |
| use ranking loss instead of distance estimator loss (ablation) | PolyD/dual · SIFT1M | recall@100: 0.987 -> 0.960 | Hurt |
| use ranking loss instead of distance estimator loss (ablation) | PolyD/binary · FYCNN1M | recall@100: 0.281 -> 0.222 | Hurt |
| add Hamming filtering (dual) on top of polysemous codes versus binary only | PolyD · SIFT1M | recall@1: 0.107 -> 0.441 | Improved |
| use polysemous dual instead of exhaustive PQ ADC | PQ (ADC) · SIFT1M | query time (ms): 9.01 -> 2.53 | Improved |
| use polysemous dual instead of exhaustive PQ ADC | PQ (ADC) · SIFT1M | recall@1: 0.442 -> 0.441 | No clear effect |
| replace disidx filtering with polysemous dual | PQ/disidx · SIFT1M | recall@1: 0.071 -> 0.441 | Improved |
| add PolyD+ADC filtering to IMI* | IMI* 4096^2 16/30k · BIGANN (16 bytes) | recall@1: 0.238 -> 0.217 | Hurt |
| add PolyD+ADC filtering to IMI* | IMI* 4096^2 16/30k · BIGANN (16 bytes, batch) | query time (ms): 0.75 -> 0.38 | Improved |
Closest papers by meaning
- Distance Encoded Product Quantization for Approximate K-Nearest Neighbor Search in High-Dimensional Space (TPAMI 2018 · similarity 0.91)
- Composite Quantization (TPAMI 2018 · similarity 0.89)
- Approximate Search with Quantized Sparse Representations (ECCV 2016 · similarity 0.88)
- AnnArbor: Approximate Nearest Neighbors Using Arborescence Coding (ICCV 2017 · similarity 0.87)
- Product Quantizer Aware Inverted Index for Scalable Nearest Neighbor Search (ICCV 2021 · similarity 0.87)
- 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.
