Research map / Architectures and efficiency / Quantization
AutoQ: Automated Kernel-Wise Neural Network Quantization
Qian Lou, Feng Guo, Minje Kim, Lantao Liu, Lei Jiang.
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 |
|---|---|---|---|
| Assign a separate bitwidth to each weight kernel instead of one per layer | AutoQ, resource-constrained quantization · ImageNet, ResNet-18/ResNet-50/SqueezeNetV1/MobileNetV2 | top-1 error (%) | Improved |
| Assign a separate bitwidth to each weight kernel instead of one per layer | AutoQ, resource-constrained quantization · ImageNet, ResNet-18/ResNet-50/SqueezeNetV1/MobileNetV2 | inference latency (ms) | Improved |
| Assign a separate bitwidth to each weight kernel instead of one per layer | AutoQ, accuracy-guaranteed quantization · ImageNet, ResNet-18/ResNet-50/SqueezeNetV1/MobileNetV2 | inference latency (ms): 42.2% | Improved |
| Use kernel-wise rather than layer-wise quantization | AutoQ, accuracy-guaranteed quantization · ImageNet, ResNet-18/ResNet-50/SqueezeNetV1/MobileNetV2 | top-1 accuracy: -0.1% | Mixed |
| Use hierarchical DRL with HLC goals and LLC kernel actions rather than one-level DDPG | Kernel-wise quantization search on MobileNetV2 · MobileNetV2 | search reward (inference accuracy): ~70% accuracy after only 200 episodes; 20% inference accuracy until 250 episodes | Improved |
| Include extrinsic environment reward in the intrinsic reward, in addition to HLC-goal completion | AutoQ LLC (relative to HIRO intrinsic reward) · MobileNetV2 kernel-wise search | search reward (inference accuracy): extra 200 episodes for HIRO to reach only 60% accuracy | Improved |
| Use kernel-wise rather than layer-wise quantization | AutoQ on the spatial BitFunsion accelerator · CNN models evaluated on BitFunsion | inference latency: 39.04% | Improved |
| Use kernel-wise rather than layer-wise quantization | AutoQ on the spatial BitFunsion accelerator · CNN models evaluated on BitFunsion | inference energy: 33.34% | Improved |
| Quantize weight sub-kernels independently instead of whole kernels (ablation) | AutoQ on the spatial BitFunsion accelerator · CNN models evaluated on BitFunsion | inference latency | No clear effect |
Closest papers by meaning
- Incremental Network Quantization: Towards Lossless CNNs with Low-precision Weights (ICLR 2017 · similarity 0.89)
- HAQ: Hardware-Aware Automated Quantization With Mixed Precision (CVPR 2019 · similarity 0.88)
- DBQ: A Differentiable Branch Quantizer for Lightweight Deep Neural Networks (ECCV 2020 · similarity 0.88)
- Learning to Quantize Deep Networks by Optimizing Quantization Intervals With Task Loss (CVPR 2019 · similarity 0.88)
- JAQ: Joint Efficient Architecture Design and Low-Bit Quantization with Hardware-Software Co-Exploration (AAAI 2025 · similarity 0.88)
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 2024 · 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.
