Research map / Architectures and efficiency / Quantization
Norm-Explicit Quantization: Improving Vector Quantization for Maximum Inner Product Search
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Explicitly quantize relative norm with a dedicated codebook while quantizing direction with the existing VQ method | PQ, OPQ, RQ, and AQ · Netflix, Yahoo!Music, ImageNet, and SIFT100M | recall-item performance | Helped |
| Explicitly quantize norm with a dedicated codebook | RQ · Yahoo!Music | norm error: 1.51×10^{-2} -> 1.1×10^{-3} (8 codebooks); 6.47×10^{-3} -> 1.1×10^{-3} (16 codebooks) | Helped |
| Use more norm codebooks, leaving fewer codebooks for direction (ablation) | NEQ · Datasets evaluated with K=256 | recall-item performance | Hurt |
| Normalize items to unit direction vectors before training direction codebooks, then quantize relative norm (ablation) | NEQ · MIPS datasets | recall-item performance | Helped |
| Train vector codebooks directly on original items instead of normalized direction vectors (ablation) | NEQ · MIPS datasets | recall-item performance | Hurt |
| Replace original RQ with NE-RQ using one norm codebook (ablation) | RQ · Netflix, Yahoo!Music, ImageNet, and SIFT100M | quantization error | Hurt |
| Replace original RQ with NE-RQ using one norm codebook (ablation) | RQ · Netflix, Yahoo!Music, ImageNet, and SIFT100M | norm error | Helped |
| Use NE-RQ instead of RQ (ablation) | RQ · SIFT100M, varying number of codebooks and target k | recall-item performance | Helped |
| Use NEQ-PQ and NEQ-OPQ instead of PQ and OPQ (ablation) | PQ and OPQ · ImageNet, across different codebook-learning runs | recall variability | Helped |
| Use NEQ-RQ instead of RQ (ablation) | RQ · ImageNet, across different codebook-learning runs | recall variability | No clear effect |
| Use NE-PQ instead of QUIP-cov(x) | MIPS quantization · ImageNet | recall-item performance | Helped |
| Use NE-RQ with two codebooks for candidate generation instead of ip-NSW | MIPS candidate generation · ImageNet | recall at equal query processing time | Helped |
Closest papers by meaning
- Qinco2: Vector Compression and Search with Improved Implicit Neural Codebooks (ICLR 2025 · similarity 0.88)
- Residual Quantization with Implicit Neural Codebooks (ICML 2024 · similarity 0.87)
- LG-VQ: Language-Guided Codebook Learning (NeurIPS 2024 · similarity 0.87)
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.87)
- NSNQuant: A Double Normalization Approach for Calibration-Free Low-Bit Vector Quantization of KV Cache (NeurIPS 2025 · similarity 0.87)
- QuIP$\#$: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks (ICML 2024 · 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.