Research map / Architectures and efficiency / Quantization
Optimizing Information Theory Based Bitwise Bottlenecks for Efficient Mixed-Precision Activation Quantization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 Bitwise Bottleneck instead of DSQ with 5-bit activations | ResNet50 · ImageNet | Top-1 accuracy (%): 75.2 -> 75.7 | Helped |
| Use Bitwise Bottleneck instead of DSQ with 4-bit activations | ResNet50 · ImageNet | Top-1 accuracy (%): 73.8 -> 74.8 | Helped |
| Use Bitwise Bottleneck instead of rounding-based quantization | ResNet50 activation quantization · MNIST | PSNR (dB): 42.02±0.54 -> 47.30±0.51 | Helped |
| Use Bitwise Bottleneck instead of sensitivity-based quantization | ResNet50 activation quantization · MNIST | PSNR (dB): 34.94±0.84 -> 47.30±0.51 | Helped |
| Reduce the activation code rate to fewer than four bits (ablation) | Bitwise Bottleneck · MNIST and CIFAR10 | classification accuracy | Mixed |
| Use more than three activation bits rather than lower precision (ablation) | Bitwise Bottleneck · MNIST and CIFAR10 | classification accuracy | Helped |
| Increase the accepted PSNR-loss threshold (ablation) | Bitwise Bottleneck · CIFAR10 | classification accuracy: less than 1% | Hurt |
| Compress activations to fewer than 6 bits on average (ablation) | Bitwise Bottleneck · CIFAR10 | classification accuracy | Hurt |
| Use Bitwise Bottleneck rather than standard rounding-based quantization | ResNet50 activation quantization · MNIST | PSNR (dB): 1.98 dB to 12.44 dB higher PSNR | Helped |
Closest papers by meaning
- Don’t Waste Your Bits! Squeeze Activations and Gradients for Deep Neural Networks via TinyScript (ICML 2020 · similarity 0.87)
- Few-bit Backward: Quantized Gradients of Activation Functions for Memory Footprint Reduction (ICML 2023 · similarity 0.87)
- Information Bottleneck: Exact Analysis of (Quantized) Neural Networks (ICLR 2022 · similarity 0.87)
- Quantization Networks (CVPR 2019 · similarity 0.86)
- Information Bottleneck Analysis of Deep Neural Networks via Lossy Compression (ICLR 2024 · similarity 0.86)
- Approaching Rate-Distortion Limits in Neural Compression with Lattice Transform Coding (ICLR 2025 · 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.