Research map / Optimization / Implicit bias and generalization
Understanding Straight-Through Estimator in Training Activation Quantized Neural Nets
Penghang Yin, Jiancheng Lyu, Shuai Zhang, Stanley Osher, Yingyong Qi, Jack Xin
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 |
|---|---|---|---|
| Use clipped ReLU STE instead of identity STE (ablation) | LeNet5 · MNIST 2-bit activations | validation accuracy: 98.49 -> 99.23 | Improved |
| Use clipped ReLU STE instead of identity STE (ablation) | LeNet5 · MNIST 4-bit activations | validation accuracy: 98.98 -> 99.24 | Improved |
| Use clipped ReLU STE instead of vanilla ReLU STE (ablation) | LeNet5 · MNIST 2-bit activations | validation accuracy: 99.24 -> 99.23 | No clear effect |
| Use clipped ReLU STE instead of vanilla ReLU STE (ablation) | LeNet5 · MNIST 4-bit activations | validation accuracy: 99.32 -> 99.24 | Hurt |
| Use clipped ReLU STE instead of identity STE (ablation) | VGG11 · CIFAR-10 2-bit activations | validation accuracy: 86.58 -> 90.92 | Improved |
| Use clipped ReLU STE instead of vanilla ReLU STE (ablation) | VGG11 · CIFAR-10 2-bit activations | validation accuracy: 88.69 -> 90.92 | Improved |
| Use clipped ReLU STE instead of identity STE (ablation) | VGG11 · CIFAR-10 4-bit activations | validation accuracy: 90.19 -> 92.08 | Improved |
| Use clipped ReLU STE instead of vanilla ReLU STE (ablation) | VGG11 · CIFAR-10 4-bit activations | validation accuracy: 92.01 -> 92.08 | Improved |
| Use clipped ReLU STE instead of identity STE (ablation) | ResNet20 · CIFAR-10 2-bit activations | validation accuracy: 46.52 -> 88.39 | Improved |
| Use clipped ReLU STE instead of vanilla ReLU STE (ablation) | ResNet20 · CIFAR-10 2-bit activations | validation accuracy: 48.05 -> 88.39 | Improved |
| Use clipped ReLU STE instead of identity STE (ablation) | ResNet20 · CIFAR-10 4-bit activations | validation accuracy: 54.16 -> 91.24 | Improved |
| Use clipped ReLU STE instead of vanilla ReLU STE (ablation) | ResNet20 · CIFAR-10 4-bit activations | validation accuracy: 86.59 -> 91.24 | Improved |
Closest papers by meaning
- High-Dimensional Learning Dynamics of Quantized Models with Straight-Through Estimator (ICML 2026 · similarity 0.90)
- Improving the Straight-Through Estimator with Zeroth-Order Information (NeurIPS 2025 · similarity 0.89)
- Network Quantization With Element-Wise Gradient Scaling (CVPR 2021 · similarity 0.88)
- Straightening Out the Straight-Through Estimator: Overcoming Optimization Challenges in Vector Quantized Networks (ICML 2023 · similarity 0.87)
- Estimator Meets Equilibrium Perspective: A Rectified Straight Through Estimator for Binary Neural Networks Training (ICCV 2023 · similarity 0.87)
- Straight-Through Meets Sparse Recovery: the Support Exploration Algorithm (ICML 2024 · similarity 0.85)
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.
