Research map / Label-efficient and robust learning / Self-supervised and contrastive
Synergistic Self-Supervised and Quantization Learning
Yun-Hao Cao, Peiqin Sun, Yechang Huang, Jianxin Wu, Shuchang Zhou
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 |
|---|---|---|---|
| quantize only the target branch (no gradient to quantized branch) (ablation) | SimSiam on CIFAR-10 · CIFAR-10 linear eval ResNet-34, average | avg accuracy: 79.2 -> 78.0 | Hurt |
| also quantize the prediction branch and update it with gradients (ablation) | SSQL · CIFAR-10 linear eval ResNet-34, average | avg accuracy: 78.0 -> 84.5 | Improved |
| randomly select bit-widths instead of a single 4-bit setting (ablation) | SSQL · CIFAR-10 linear eval ResNet-34, average | avg accuracy: 33.4 -> 83.9 | Improved |
| randomly select bit-widths instead of a single 6-bit setting (ablation) | SSQL · CIFAR-10 linear eval ResNet-34, average | avg accuracy: 79.8 -> 83.9 | Improved |
| use narrower low-bit range 2~8/4~8 instead of 4~16 bits (ablation) | SSQL · CIFAR-10 linear eval ResNet-34, 2w4a | accuracy: 65.2 -> 81.5 | Mixed |
| add auxiliary SimSiam loss (ablation) | SSQL · CIFAR-10 linear eval ResNet-34, average | avg accuracy: 86.3 -> 88.3 | Improved |
| also quantize the target branch with 2~8 bit range (ablation) | SSQL · CIFAR-10 linear eval ResNet-34, average | avg accuracy: 86.3 -> 85.8 | Hurt |
| use SSQL pretraining instead of SimSiam | ResNet-18 · CIFAR-10 linear eval (3w3a) | accuracy: 66.0 -> 85.6 | Improved |
| use SSQL pretraining instead of SimSiam | ResNet-18 · ImageNet linear eval (FP) | accuracy: 55.0 -> 57.6 | Improved |
| use SSQL pretraining instead of SimSiam | ResNet-50 · ImageNet linear eval (4w4a) | accuracy: 52.4 -> 63.0 | Improved |
| add SSQL to BYOL | ResNet-18 · CIFAR-10 linear eval (2w4a) | accuracy: 63.3 -> 85.7 | Improved |
| add SSQL to MoCov3 | ViT-Small · CIFAR-10 linear eval (4w4a) | accuracy: 82.0 -> 86.9 | Improved |
Closest papers by meaning
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.87)
- RSQ: Learning from Important Tokens Leads to Better Quantized LLMs (NeurIPS 2026 · similarity 0.87)
- GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling (NeurIPS 2026 · similarity 0.86)
- SynQ: Accurate Zero-shot Quantization by Synthesis-aware Fine-tuning (ICLR 2025 · similarity 0.86)
- Q-Palette: Fractional-Bit Quantizers Toward Optimal Bit Allocation for Efficient LLM Deployment (NeurIPS 2025 · similarity 0.86)
- SliM-LLM: Salience-Driven Mixed-Precision Quantization for Large Language Models (ICML 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.
