Research map / Architectures and efficiency / Quantization
LSQ++: Lower running time and higher recall in multi-codebook quantization
Julieta Martinez, Shobhit Zakhmi, Holger H. Hoos, James J. Little
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 |
|---|---|---|---|
| replace conjugate gradient codebook update with direct regularized Cholesky solution (ablation) | LSQ (Julia, C++) · SIFT1M 64 bits | time per iteration: 14.2 s -> 5.6 s | Improved |
| replace conjugate gradient codebook update with direct regularized Cholesky solution (ablation) | LSQ (Julia, CUDA) · SIFT1M 128 bits | time per iteration: 20.3 s -> 4.3 s | Improved |
| replace conjugate gradient codebook update with direct regularized Cholesky solution (ablation) | LSQ (Julia, CUDA) · Deep1M 64 bits | time per iteration: 3.3 s -> 1.0 s | Improved |
| add stochastic relaxation SR-D (noise in codebooks) to LSQ | LSQ · SIFT1M 64 bits | Recall@1: 0.340 -> 0.346 | Improved |
| add SR-C (noise added to data) to LSQ at 128 bits | LSQ · MNIST, LabelMe22K, SIFT1M 128 bits | Recall@1 | Hurt |
| add SR-C to LSQ at 64 bits | LSQ · 64-bit datasets | Recall@1 | Mixed |
| double LSQ++ iterations from 25 to 50 (ablation) | LSQ++ · SIFT1M 64 bits | Recall@1: 0.346 -> 0.348 | Improved |
| increase LSQ++ to 100 iterations with 128 encoding iterations (ablation) | LSQ++ · SIFT1M 64 bits | Recall@1: 0.348 -> 0.351 | Improved |
| double base encoding ILS iterations from 128 to 256 (ablation) | LSQ++ · SIFT1M 64 bits | Recall@1: 0.351 -> 0.353 | Improved |
| train CQ on the learn set instead of the base set (ablation) | CQ · SIFT1M 64 bits | Recall@1: 0.290 -> 0.162 | Hurt |
| use norm codebook so RVQ and ERVQ account for query time | RVQ, ERVQ · MNIST, LabelMe22K, SIFT1M | Recall@1 | Hurt |
Closest papers by meaning
- MBQ: Modality-Balanced Quantization for Large Vision-Language Models (CVPR 2025 · similarity 0.88)
- GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling (NeurIPS 2026 · similarity 0.87)
- LG-VQ: Language-Guided Codebook Learning (NeurIPS 2024 · similarity 0.86)
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.86)
- Extreme Compression of Large Language Models via Additive Quantization (ICML 2024 · similarity 0.86)
- Q-Palette: Fractional-Bit Quantizers Toward Optimal Bit Allocation for Efficient LLM Deployment (NeurIPS 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.
