Research map / Architectures and efficiency / Quantization
Post-Training Sparsity-Aware Quantization
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 |
|---|---|---|---|
| Use five rather than two bit-window placement options (ablation) | SPARQ 4-bit · ILSVRC-2012, ResNet-18 | top-1 accuracy degradation: -1.37% -> -0.07% | Helped |
| Add rounding within the selected bit window (ablation) | SPARQ 5opt · ILSVRC-2012, ResNet-18 | top-1 accuracy degradation: -0.11% -> -0.07% | Helped |
| Add vSPARQ pairing of activations (ablation) | SPARQ 3-bit · ILSVRC-2012, ResNet-18 | top-1 accuracy degradation: -0.51% -> -0.21% | Helped |
| Add vSPARQ pairing of activations (ablation) | SPARQ 2-bit · ILSVRC-2012, ResNet-18 | top-1 accuracy degradation: -2.57% -> -1.64% | Helped |
| Add vSPARQ pairing of activations (ablation) | SPARQ 4-bit 5opt · ILSVRC-2012, ResNet-18 | top-1 accuracy degradation: -0.11% -> -0.11% | No clear effect |
| Reduce window options from five to three (ablation) | SPARQ 4-bit · ILSVRC-2012, ResNet-18 | top-1 accuracy degradation: -0.07% -> -0.14% | Hurt |
| Reduce window options from three to two (ablation) | SPARQ 4-bit · ILSVRC-2012, ResNet-18 | top-1 accuracy degradation: -0.14% -> -1.37% | Hurt |
| Use 2-bit rather than 3-bit activations (ablation) | SPARQ · ILSVRC-2012, ResNet-18 | top-1 accuracy degradation: -0.21% -> -1.64% | Hurt |
| Apply SPARQ to a 2:4-pruned model (ablation) | SPARQ 4-bit 5opt · ILSVRC-2012, ResNet-18 | top-1 accuracy degradation: -0.07% -> -0.13% | Hurt |
| Use the SPARQ 3opt processing element instead of conventional 8b-8b | Systolic Array PE · Systolic Array | relative hardware area: 1.00 -> 0.61 | Helped |
| Use the SPARQ 3opt processing element instead of conventional 8b-8b | Tensor Core PE · Tensor Core | relative hardware area: 1.00 -> 0.66 | Helped |
Closest papers by meaning
- BSQ: Exploring Bit-Level Sparsity for Mixed-Precision Neural Network Quantization (ICLR 2021 · similarity 0.91)
- PTMQ: Post-training Multi-Bit Quantization of Neural Networks (AAAI 2024 · similarity 0.90)
- Effective Interplay between Sparsity and Quantization: From Theory to Practice (ICLR 2025 · similarity 0.90)
- Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · similarity 0.89)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.88)
- Reg-PTQ: Regression-specialized Post-training Quantization for Fully Quantized Object Detector (CVPR 2024 · similarity 0.88)
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.