Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
SynQ: Accurate Zero-shot Quantization by Synthesis-aware Fine-tuning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SynQ, a zero-shot quantization fine-tuning framework that low-pass filters synthetic samples, aligns Grad-CAM maps between teacher and quantized model, and drops hard-label CE for difficult samples.
Problem
Zero-shot quantization (no real training data) of pre-trained classifiers suffers from high-frequency noise in synthetic data, off-target predictions of the quantized model, and misguidance from erroneous hard labels on difficult synthetic samples.
Approach
Three plug-in fixes for synthetic-data fine-tuning: a Gaussian frequency-domain low-pass filter on samples, a Grad-CAM alignment loss between the pre-trained and quantized models, and using only KL soft-label loss for samples whose difficulty exceeds a threshold.
Key results
- On ResNet-18 ImageNet at W3A3, SynQ reaches 52.02% vs. 50.28% for TexQ (about +1.74 points), the largest reported gain over prior methods.
- Applied as a plug-in, SynQ improves GDFQ, Qimera, AdaDFQ, IntraQ, HAST and TexQ, with gains up to 31.17 points (Qimera W3A3).
- On ViT quantization with PSAQ-ViT at W4A8, average accuracy rises from 70.61 to 71.19.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add low-pass filtering, CAM alignment, and soft labels for difficult samples | ZSQ quantization · ImageNet, W3A3, ResNet-18 | top-1 accuracy (%): 50.28 -> 52.02 | Helped |
| add low-pass filter (I1) (ablation) | baseline ZSQ · ImageNet, W3A3, ResNet-18 | accuracy (%): 43.63 -> 49.43 | Helped |
| add CAM alignment (I2) (ablation) | baseline ZSQ · ImageNet, W3A3, ResNet-18 | accuracy (%): 43.63 -> 48.26 | Helped |
| add soft labels for difficult samples (I3) (ablation) | baseline ZSQ · ImageNet, W3A3, ResNet-18 | accuracy (%): 43.63 -> 46.42 | Helped |
| add low-pass filter, CAM alignment, and soft labels for difficult samples (ablation) | baseline ZSQ · ImageNet, W3A3, ResNet-18 | accuracy (%): 43.63 -> 52.02 | Helped |
| use CAM alignment instead of feature alignment (ablation) | ZSQ alignment with low-pass filter and soft labels · ImageNet, W3A3, ResNet-18 | accuracy (%): 51.20 ± 0.30 -> 52.02 ± 0.34 | Helped |
| apply SynQ to PSAQ-ViT | PSAQ-ViT · ImageNet ViT models, W4A8 average | top-1 accuracy (%): 70.61 -> 71.19 | Helped |
| increase difficulty threshold above 0.5 | SynQ · ImageNet, W3A3, ResNet-18 | accuracy | Hurt |
Surprising findings
- Qimera at W3A3 on ImageNet collapses to 1.17%, and SynQ recovers it to 32.34%, indicating the components are especially helpful where baselines fail.
Limitations
- Gains on the noise-free and low-bit regime depend on the baseline; improvements are small (0.12-0.63 points) for some strong baselines like TexQ at W4A4.
- Robustness to different noise types varies; the filter helps less for some noise types (e.g., speckle).
- Evaluated only on image classification with CNNs and ViTs; object detection and diffusion models are left for future work.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 ImageNet
Benchmarks: CIFAR-10 CIFAR-100 ImageNet
Metrics: top-1 accuracy standard deviation over 5 seeds KL divergence between saliency maps
Baselines: GDFQ ARC Qimera ARC + AIT IntraQ AdaSG AdaDFQ HAST TexQ PLF PSAQ-ViT Genie
Builds on: GDFQ HAST TexQ IntraQ Grad-CAM
Closest papers by meaning
- ZeroQ: A Novel Zero Shot Quantization Framework (CVPR 2020 · similarity 0.92)
- Zero-Shot Adversarial Quantization (CVPR 2021 · similarity 0.91)
- Hard Sample Matters a Lot in Zero-Shot Quantization (CVPR 2023 · similarity 0.91)
- Sharpness-Aware Data Generation for Zero-shot Quantization (ICML 2024 · similarity 0.89)
- Zero-Shot Quantization via Weight-Space Arithmetic (NeurIPS 2026 · similarity 0.88)
- GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling (NeurIPS 2026 · 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.