Research map / Graph learning / GNN architectures
Degree-Quant: Quantization-Aware Training for Graph Neural Networks
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 |
|---|---|---|---|
| Apply Degree-Quant at W4A4 instead of standard QAT | GIN · Cora | accuracy: 42.5±4.5 -> 69.9±3.4 | Helped |
| Apply Degree-Quant at W4A4 instead of standard QAT | GIN · Citeseer | accuracy: 18.6±2.9 -> 60.8±2.1 | Helped |
| Apply Degree-Quant at W4A4 instead of noisy QAT | GIN · REDDIT-BINARY | accuracy: 58.0±6.3 -> 81.3±4.4 | Helped |
| Apply Degree-Quant at W8A8 instead of standard QAT | GIN · ZINC | loss: 0.386±0.025 -> 0.357±0.014 | Helped |
| Apply Degree-Quant at W8A8 instead of FP32 reference | GIN · REDDIT-BINARY | accuracy: 92.2±2.3 -> 91.8±2.3 | No clear effect |
| Use INT8 arithmetic instead of FP32 | GAT inference on CPU · Reddit | speedup: 4.7× | Helped |
| Add degree-based stochastic protective masking (ablation) | INT8 GIN · Cora | accuracy: 75.6±1.2 -> 78.9±1.2 | Helped |
| Add degree-based stochastic protective masking (ablation) | INT4 GIN · Cora | accuracy: 42.5±4.5 -> 71.2±2.9 | Helped |
| Add percentile ranges to DQ masking (ablation) | INT4 GIN · Cora | accuracy: 71.2±2.9 -> 69.9±3.4 | Hurt |
| Replace degree-aware protective masking with noisy QAT while retaining percentiles (ablation) | INT4 GIN · Cora | accuracy: 63.5±2.0 versus 69.9±3.4 | Hurt |
| Use momentum rather than min/max range tracking with vanilla STE (ablation) | W8A8 GCN · Cora | accuracy: 81.0±0.7 -> 42.3±11.1 | Hurt |
| Add gradient clipping to momentum-based range tracking (ablation) | W8A8 GCN · Cora | accuracy: 42.3±11.1 -> 66.9±18.2 | Helped |
Closest papers by meaning
- $\rm A^2Q$: Aggregation-Aware Quantization for Graph Neural Networks (ICLR 2023 · similarity 0.93)
- VQ-GNN: A Universal Framework to Scale up Graph Neural Networks using Vector Quantization (NeurIPS 2021 · similarity 0.89)
- Generalizing CNNs to graphs with learnable neighborhood quantization (NeurIPS 2024 · similarity 0.89)
- Linear Symmetric Quantization of Neural Networks for Low-precision Integer Hardware (ICLR 2020 · similarity 0.89)
- Towards Efficient Verification of Quantized Neural Networks (AAAI 2024 · similarity 0.87)
- A Manifold Perspective on the Statistical Generalization of Graph Neural Networks (ICML 2025 · similarity 0.87)
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.