Research map / Architectures and efficiency / Quantization
Simple strategies for recovering inner products from coarsely quantized random projections
Ping Li, Martin Slawski
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 |
|---|---|---|---|
| normalize projected vectors (normalized estimator instead of linear estimator) | full-precision random projection cosine estimation · rho=0.99 | asymptotic MSE ratio: 5000 | Improved |
| normalize projected vectors at low similarity | full-precision random projection cosine estimation · rho=0.5 | asymptotic MSE ratio: 2.2 | Improved |
| use MLE instead of normalized estimator | full-precision cosine estimation · general | MSE: at most by a factor of 2 | Improved |
| raise bit depth from 3 to 4 in linear code-based estimator | linear estimator · rho=1 squared bias bound | squared bias bound: 7.2 · 10−3 -> 4.5 · 10−4 | Improved |
| raise bit depth from 2 to 3 | linear estimator · rho=1 squared bias bound | squared bias bound: 1.2 · 10−1 -> 7.2 · 10−3 | Improved |
| quantize to 4 bits instead of full precision with k=1000 | linear estimator · MSE vs full precision | MSE: at most a factor of 1.45 | Hurt |
| add normalization to quantized linear estimator (ablation) | quantized code-based estimator · bias reduction | squared bias | Improved |
| collision-based (b=2, doubled k) instead of normalized (b=4) | quantized cosine estimation at equal bits · high similarity rho | MSE: crossover rho from about 0.55 to 0.95 | Mixed |
| increase b for collision-based estimator | collision-based estimator · MSE vs rho | MSE | Mixed |
| increase b for code-based estimators | normalized estimator · MSE over whole rho range | MSE | Improved |
| increase k beyond order of squared bias | quantized linear estimator · MSE, large rho | MSE | No clear effect |
| use normalized b=4 instead of collision b=2 with doubled k | SVM on arcene, Dexter, farm, PEMS · UCI classification | test accuracy | Improved |
Closest papers by meaning
- Quantized Random Projections and Non-Linear Estimation of Cosine Similarity (NeurIPS 2016 · similarity 0.90)
- Random Projections with Asymmetric Quantization (NeurIPS 2019 · similarity 0.88)
- Generalization Error Analysis of Quantized Compressive Learning (NeurIPS 2019 · similarity 0.87)
- Quantization Algorithms for Random Fourier Features (ICML 2021 · similarity 0.87)
- TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate (ICLR 2026 · similarity 0.86)
- QuIP$\#$: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks (ICML 2024 · 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.
