Research map / Architectures and efficiency / Quantization
Is Bin Generation Indispensable? A Bin-Generation-Free Dataset Quantization via Semantic Perspective
Maijie Deng, Yuhua Li, Yixiong Zou, Yao Wu, Chenru Ma
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 bin generation with KNN neighbor identification | BGFDQ (w/o SPD) · CIFAR-10 runtime (rho=20%) | total runtime (s): 1248 -> 1052 | Improved |
| add semantic-shift patch dropping (SPD) (ablation) | BGFDQ · CIFAR-100 (rho=20%) | validation accuracy: 62.8 -> 66.0 | Improved |
| add semantic-shift patch dropping (SPD) (ablation) | BGFDQ · CIFAR-10 runtime (rho=20%) | total runtime (s): 1052 -> 2465 | Hurt |
| use neighbor-aware coreset selection instead of bin generation + random selection (ablation) | coreset selection · CIFAR-10, 5K samples | validation accuracy: 75.9 -> 76.4 | Improved |
| add bin generation before random selection (ablation) | random selection · CIFAR-10 and CIFAR-100 | validation accuracy: 0.1% average improvement | No clear effect |
| raise vanilla patch drop ratio from 25% to 40% (ablation) | patch dropping · CIFAR-10 random subsets | accuracy change vs unquantized | Hurt |
| use adaptive SPD (avg 40% drop) instead of vanilla 40% patch dropping (ablation) | patch dropping · CIFAR-10 random subsets | accuracy change vs unquantized | Improved |
| use BGFDQ instead of DQ coresets for ViT training | coreset · CIFAR-100 ViT (rho=20%) | validation accuracy: 30.8 -> 35.7 | Improved |
| use BGFDQ on Fakeddit where bin-generation methods run out of memory | coreset selection · Fakeddit (rho=1%) | validation accuracy: 83.3 -> 83.8 | Improved |
| vary neighbor count K (ablation) | neighbor identification · CIFAR-10 | validation accuracy | No clear effect |
Closest papers by meaning
- Adaptive Dataset Quantization (AAAI 2025 · similarity 0.88)
- BiDM: Pushing the Limit of Quantization for Diffusion Models (NeurIPS 2024 · similarity 0.85)
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 2024 · similarity 0.85)
- ARB-LLM: Alternating Refined Binarizations for Large Language Models (ICLR 2025 · similarity 0.84)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.83)
- Rethinking Dataset Distillation for Classification: Do Distilled Sets Outperform Coresets? (NeurIPS 2026 · similarity 0.83)
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.
