Research map / Architectures and efficiency / Quantization
PROFIT: A Novel Training Method for sub-4-bit MobileNet Models
Eunhyeok Park, Sungjoo Yoo
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 |
|---|---|---|---|
| add PROFIT (with TS and PG) (ablation) | PACT, 4-bit MobileNet-v3 · ImageNet val, 4-bit MobileNet-v3 ablation | top-1 accuracy [%]: 70.160 -> 72.948 | Improved |
| add PROFIT (with TS and PG) (ablation) | DuQ, 4-bit MobileNet-v3 · ImageNet val, 4-bit MobileNet-v3 ablation | top-1 accuracy [%]: 71.466 -> 73.812 | Improved |
| add PROFIT alone (ablation) | DuQ, 4-bit MobileNet-v3 · ImageNet val, 4-bit MobileNet-v3 ablation | top-1 accuracy [%]: 69.504 -> 72.260 | Improved |
| use DuQ instead of PACT (ablation) | 4-bit MobileNet-v3 · ImageNet val, 4-bit MobileNet-v3 ablation | top-1 accuracy [%]: 67.978 -> 69.504 | Improved |
| use DuQ+PROFIT instead of PACT under TS and PG (ablation) | 4-bit MobileNet-v3 · ImageNet val, 4-bit MobileNet-v3 ablation | top-1 accuracy [%]: 70.160 -> 73.812 | Improved |
| add teacher-student distillation (ablation) | DuQ, 4-bit MobileNet-v3 · ImageNet val, 4-bit MobileNet-v3 ablation | top-1 accuracy [%]: 69.504 -> 71.006 | Improved |
| add progressive quantization (ablation) | DuQ+TS, 4-bit MobileNet-v3 · ImageNet val, 4-bit MobileNet-v3 ablation | top-1 accuracy [%]: 71.006 -> 71.466 | Improved |
| use full PROFIT instead of only the final normalization-layer stage (ablation) | DuQ+TS+PG · ImageNet val, 4-bit MobileNet-v3 ablation | top-1 accuracy [%]: 72.826 -> 73.812 | Improved |
| use negative padding instead of zero padding with symmetric quantization (ablation) | DuQ, 4-bit MobileNet-v3 · ImageNet val, 4-bit MobileNet-v3 ablation | top-1 accuracy [%]: 68.480 -> 69.504 | Improved |
| use negative padding instead of zero padding with non-negative quantization (ablation) | DuQ, 4-bit MobileNet-v3 · ImageNet val, 4-bit MobileNet-v3 ablation | top-1 accuracy [%]: 68.724 -> 69.504 | Improved |
| use QIL activation quantization (ablation) | 4-bit MobileNet-v3 · ImageNet val, 4-bit MobileNet-v3 ablation | top-1 accuracy [%]: fail | Hurt |
| add PROFIT to PACT at 3 bits (ablation) | PACT, MobileNet-v3 3-bit · ImageNet val, 4-bit MobileNet-v3 ablation (3-bit) | top-1 accuracy [%]: 57.086 -> 66.458 | Improved |
Closest papers by meaning
- BASQ: Branch-Wise Activation-Clipping Search Quantization for Sub-4-Bit Neural Networks (ECCV 2022 · similarity 0.88)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.87)
- Overcoming Oscillations in Quantization-Aware Training (ICML 2022 · similarity 0.87)
- PackQViT: Faster Sub-8-bit Vision Transformers via Full and Packed Quantization on the Mobile (NeurIPS 2023 · similarity 0.87)
- MetaMix: Meta-State Precision Searcher for Mixed-Precision Activation Quantization (AAAI 2024 · similarity 0.86)
- Improving Low-Precision Network Quantization via Bin Regularization (ICCV 2021 · 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.
