Research map / Architectures and efficiency / Quantization
LQ-Nets: Learned Quantization for Highly Accurate and Compact Deep Neural Networks
Dongqing Zhang, Jiaolong Yang, Dongqiangzi Ye, Gang Hua
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 |
|---|---|---|---|
| optimize quantizers with QEM instead of backprop (ablation) | LQ-Nets ResNet-20 · CIFAR-10 (2/32) | accuracy: 90.0 -> 91.8 | Improved |
| optimize quantizers with QEM instead of backprop (ablation) | LQ-Nets ResNet-20 · CIFAR-10 (2/2) | accuracy: 88.2 -> 90.2 | Improved |
| increase QEM iterations T from 1 to 4 (ablation) | LQ-Nets ResNet-20 · CIFAR-10 (2/2) | accuracy: 90.16 -> 90.25 | No clear effect |
| increase QEM iterations T from 1 to 4 (ablation) | LQ-Nets ResNet-20 · CIFAR-10 (1/2) | accuracy: 88.37 -> 88.51 | No clear effect |
| increase QEM iterations T from 1 to 4 (ablation) | LQ-Nets ResNet-20 · CIFAR-10 (3/3) | accuracy: 91.58 -> 91.44 | No clear effect |
| learn the quantizer instead of fixed half-wave Gaussian quantizer | ResNet-18 1/2 · ImageNet | top-1 accuracy: 59.6 -> 62.6 | Improved |
| learn the quantizer instead of fixed uniform quantizer | ResNet-18 1/2 vs DoReFa 1/4 · ImageNet | top-1 accuracy: 59.2 -> 62.6 | Improved |
| learned quantizer instead of HWGQ | VGG-Small 1/2 · CIFAR-10 | accuracy: 92.5 -> 93.4 | Improved |
| increase bits from 1/2 to 4/4 (ablation) | LQ-Nets ResNet-50 · ImageNet | top-1 accuracy: 68.7 -> 75.1 | Improved |
| increase activation bits 2 to 3 to 4 with 2/32 to 4/32 weights (ablation) | LQ-Nets ResNet-18 · ImageNet | top-1 accuracy: 68.0 -> 70.0 | Improved |
| quantize to 1/2 bits instead of full precision (ablation) | LQ-Nets ResNet-18 · ImageNet | top-1 accuracy: 70.3 -> 62.6 | Hurt |
| increase weight bits from 1 to 3 on CIFAR-10 (ablation) | LQ-Nets VGG-Small · CIFAR-10 | accuracy: 93.5 -> 93.8 | Improved |
Later papers that compared with this method
1 later accepted paper reports LQNets 4/32 as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Finding the Task-Optimal Low-Bit Sub-Distribution in Deep Neural Networks (ICML 2022 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- XNOR-Net 1/1 — XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks (ECCV 2016) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- LEARNED STEP SIZE QUANTIZATION (ICLR 2020 · similarity 0.91)
- Deep Neural Network Compression With Single and Multiple Level Quantization (AAAI 2018 · similarity 0.91)
- Incremental Network Quantization: Towards Lossless CNNs with Low-precision Weights (ICLR 2017 · similarity 0.91)
- SYQ: Learning Symmetric Quantization for Efficient Deep Neural Networks (CVPR 2018 · similarity 0.91)
- Explicit Loss-Error-Aware Quantization for Low-Bit Deep Neural Networks (CVPR 2018 · similarity 0.90)
- Divide and Conquer: Leveraging Intermediate Feature Representations for Quantized Training of Neural Networks (ICML 2020 · similarity 0.90)
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.
