Research map / Architectures and efficiency / Quantization
Hierarchical Channel-spatial Encoding for Communication-efficient Collaborative Learning
Qihua Zhou, Song Guo, LIU Yi, Jie Zhang, Jiewei Zhang, Tao GUO, Zhenda XU, XUN LIU, Zhihao Qu
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 |
|---|---|---|---|
| replace uniform quantization with stripe-wise group quantization (SGQ) | feature compression · mini-ImageNet ResNet34, 4-bit | accuracy (%): 36.89 -> 74.37 | Improved |
| replace product quantization with SGQ | feature compression · mini-ImageNet ResNet34, 4-bit | accuracy (%): 69.61 -> 74.37 | Improved |
| replace PQ with SGQ | feature compression · mini-ImageNet ShuffleNet-x1.0, 4-bit | accuracy (%): 70.16 -> 74.86 | Improved |
| replace PQ with SGQ | feature compression · CIFAR-10 MobileNet-V1, 4-bit | accuracy (%): 92.67 -> 93.45 | Improved |
| replace PQ with SGQ | feature compression · Fashion-MNIST VGG11, 4-bit | accuracy (%): 95.94 -> 96.57 | Improved |
| replace CLIO top-slice dropping with SGQ at same compression | feature compression · Fashion-MNIST VGG11, 4-bit | accuracy (%): 21.02 -> 96.57 | Improved |
| SGQ 4-bit compression versus FP32 training | SGQ · mini-ImageNet ResNet34 | accuracy (%): 80.31 -> 74.37 | Hurt |
| SGQ 4-bit compression versus FP32 training | SGQ · Fashion-MNIST VGG11 | accuracy (%): 97.55 -> 96.57 | Mixed |
| lower quantization bits from 8 to 4 | SGQ · system overhead | total overhead proportion (%): 43.57 -> 9.89 | Improved |
| lower quantization bits from 4 to 2 | SGQ · system overhead | total overhead proportion (%): 9.89 -> 8.14 | Improved |
| use pixel sampling via average pooling in clustering | SGQ module · system overhead, 4 bits | SGQ module overhead (%): 31.93 -> 5.51 | Improved |
| use fewer quantization bits (4,5,6) (ablation) | SGQ · accuracy across benchmarks | average accuracy | Mixed |
Closest papers by meaning
- JointSQ: Joint Sparsification-Quantization for Distributed Learning (CVPR 2024 · similarity 0.85)
- Instance-Aware Group Quantization for Vision Transformers (CVPR 2024 · similarity 0.85)
- Permute, Quantize, and Fine-Tune: Efficient Compression of Neural Networks (CVPR 2021 · similarity 0.85)
- SDP4Bit: Toward 4-bit Communication Quantization in Sharded Data Parallelism for LLM Training (NeurIPS 2024 · similarity 0.83)
- Learning Grouped Lattice Vector Quantizers for Low-Bit LLM Compression (NeurIPS 2025 · similarity 0.83)
- Network Memory Footprint Compression Through Jointly Learnable Codebooks and Mappings (ICLR 2024 · similarity 0.83)
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.
