Research map / Architectures and efficiency / Quantization
FleXOR: Trainable Fractional Quantization
Dongsoo Lee, Se Jung Kwon, Byeongwook Kim, Yongkweon Jeon, Baeseong Park, Jeongin Yun
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 FleXOR for binary weight compression at 0.8 bit/weight instead of 1-bit BWN | ResNet-32 · CIFAR-10 | accuracy: 89.49% -> 91.20% | Improved |
| Reduce FleXOR to 0.4 bit/weight | ResNet-32 · CIFAR-10 | accuracy: 91.36% -> 89.61% | Hurt |
| Use FleXOR at 0.8 bit/weight instead of BinaryRelax at 1 bit | ResNet-18 · ImageNet | Top-1 accuracy: 63.2% -> 63.8% | Improved |
| Assign layer-specific N_in values rather than N_in=12 across all layer groups (ablation) | FleXOR · CIFAR-10, ResNet-20 | accuracy: 89.16% -> 89.29% | Improved |
| Assign layer-specific N_in values rather than N_in=12 across all layer groups (ablation) | FleXOR · CIFAR-10, ResNet-20 | bits/weight: 0.60 -> 0.47 | Improved |
| Use N_tap=2 rather than randomly populated XOR rows (ablation) | FleXOR · CIFAR-10, ResNet-32 | test accuracy | Improved |
| Use sign forward propagation and tanh-based custom gradients rather than STE for XOR (ablation) | FleXOR · CIFAR-10, ResNet-32 | test accuracy | Improved |
| Increase S_tanh excessively (ablation) | FleXOR · CIFAR-10, ResNet-32 | test accuracy | Hurt |
| Use learning-rate and S_tanh warmup (ablation) | FleXOR · CIFAR-10, ResNet-32 | test accuracy | Improved |
| Use weight clipping (ablation) | FleXOR · CIFAR-10, ResNet-32 | test accuracy | No clear effect |
| Increase the number of XOR configurations from q=1 to q=2 (ablation) | FleXOR · ImageNet, ResNet-18 | Top-1 accuracy: 63.8% -> 63.8% (0.8 bit); 63.8% -> 66.2% (1.6 bit) | Mixed |
Closest papers by meaning
- Structured Compression by Weight Encryption for Unstructured Pruning and Quantization (CVPR 2020 · similarity 0.88)
- Multi-Precision Quantized Neural Networks via Encoding Decomposition of {-1,+1} (AAAI 2019 · similarity 0.85)
- Q-Palette: Fractional-Bit Quantizers Toward Optimal Bit Allocation for Efficient LLM Deployment (NeurIPS 2025 · similarity 0.83)
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 2024 · similarity 0.83)
- QuIP$\#$: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks (ICML 2024 · similarity 0.82)
- Towards Efficient Verification of Quantized Neural Networks (AAAI 2024 · similarity 0.82)
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.
