Research map / Architectures and efficiency / Quantization
Heterogeneous Bitwidth Binarization in Convolutional Neural Networks
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 1.4-bit heterogeneous weight quantization instead of homogeneous 2-bit weight quantization | AlexNet on ImageNet, weights only · ImageNet | Top-1 accuracy: 55.3% -> 55.2% | No clear effect |
| use 1.4-bit heterogeneous inputs and weights instead of 2-bit inputs and 1-bit weights | AlexNet · ImageNet | Top-1 accuracy: 51.0% -> 53.2% | Helped |
| use 1.4-bit heterogeneous inputs and 1-bit weights instead of 2-bit inputs and 1-bit weights | AlexNet · ImageNet | Top-1 accuracy: 51.0% -> 51.5% | Helped |
| use 1-bit inputs and 1.4-bit weights instead of 1-bit inputs and 1-bit weights | MobileNet · ImageNet | Top-1 accuracy: 52.9% -> 60.1% | Helped |
| use 1.4-bit inputs and 1-bit weights instead of 1-bit inputs and 1-bit weights | MobileNet · ImageNet | Top-1 accuracy: 52.9% -> 62.0% | Helped |
| use middle-out bit selection instead of top-down, bottom-up, or random selection (ablation) | 1.4-bit HBNN AlexNet-BN · ImageNet | accuracy | Helped |
| assign fewer bits to values closest to the mean rather than using an alternative sorting heuristic (ablation) | heterogeneous tensor approximation · normally distributed tensor | normalized Euclidean distance | Helped |
| vary the mix of 1-, 2-, and 3-bit parameters while keeping the same 1.4-bit average (ablation) | AlexNet HBNN · ImageNet | Top-1 accuracy: 49.4% and 49.3% | No clear effect |
| increase average bitwidth across homogeneous layer-level assignments (ablation) | four-layer convolutional model · CIFAR-10 | accuracy | Helped |
| increase average bitwidth above 1.4 bits (ablation) | AlexNet HBNN, weights and activations · ImageNet | Top-1 accuracy | No clear effect |
| reduce MobileNet ASIC bitwidth from 2 bits to 1.4 bits | MobileNet ASIC implementation · ImageNet | chip area: 297 mm2 -> 145.5 mm2 | Helped |
| reduce VGG-8 ASIC bitwidth from 2 bits to 1.2 bits | VGG-8 ASIC implementation · CIFAR-10 | Pchip: 0.38 -> 0.14 | Helped |
Closest papers by meaning
- Not All Bits have Equal Value: Heterogeneous Precisions via Trainable Noise (NeurIPS 2022 · similarity 0.87)
- Towards Effective Low-Bitwidth Convolutional Neural Networks (CVPR 2018 · similarity 0.87)
- Mixed Precision DNNs: All you need is a good parametrization (ICLR 2020 · similarity 0.86)
- BD-Net: Has Depth-Wise Convolution Ever Been Applied in Binary Neural Networks? (AAAI 2026 · similarity 0.86)
- Binarized Neural Networks (NeurIPS 2016 · similarity 0.86)
- BiBench: Benchmarking and Analyzing Network Binarization (ICML 2023 · similarity 0.86)
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.