Research map / Architectures and efficiency / Quantization
A&B BNN: Add&Bit-Operation-Only Hardware-Friendly Binary Neural Network
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes A&B BNN, a multiplication-free binary neural network built on BN-Free ReActNet by adding a removable mask layer and a power-of-2 quantized RPReLU, reaching 66.89% ImageNet top-1 with zero multiplications at inference.
Problem
Binary neural networks still contain millions of full-precision multiplications from BN, PReLU, scaling factors, and pooling, which are hardware-unfriendly and conflict with BNN's goal of logic-only inference.
Approach
Build on BN-Free binary networks, absorb the beta multiplication into a training-only mask layer that is removed at inference, constrain PReLU slopes and alpha to powers of 2 so multiplications become bit shifts, and replace average pooling division with 2-bit right shifts.
Key results
- A&B ReActNet-A reaches 66.89% top-1 on ImageNet with zero multiplications, versus 68.0% for BN-Free ReActNet-A (14.65M multiplications).
- A&B ReActNet-18 reaches 61.39% top-1 on ImageNet, 0.29% above BN-Free ReActNet-18 while removing 4.6M multiplications.
- Quantized RPReLU improves ImageNet top-1 by 1.14% over fixed-slope RLeakyReLU for ReActNet-18.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace BN-Free with A&B BNN | ReActNet-18 · ImageNet | Top-1 accuracy: 61.1% -> 61.39% | Helped |
| Replace BN-Free with A&B BNN | ReActNet-A · ImageNet | Top-1 accuracy: 68.0% -> 66.89% | Hurt |
| Replace BN-Free with A&B BNN | ReActNet-18 · CIFAR-10 | accuracy: 92.08 -> 92.30 | Helped |
| Replace BN-Free with A&B BNN | ReActNet-18 · CIFAR-100 | accuracy: 68.34 -> 69.35 | Helped |
| Replace fixed-slope RLeakyReLU (slope 2−3) with quantized RPReLU (ablation) | ReActNet-18 · ImageNet | Top-1 accuracy: 60.25 -> 61.39; 1.14% | Helped |
| Replace fixed-slope RLeakyReLU (slope 2−7) with quantized RPReLU (ablation) | ReActNet-18 · ImageNet | Top-1 accuracy: 60.48 -> 61.39 | Helped |
| Replace fixed-slope RLeakyReLU with quantized RPReLU (ablation) | ReActNet-18 · CIFAR-10 | Top-1 accuracy: 92.30 -> 91.94; 92.07 -> 91.94 | Mixed |
| Replace fixed-slope RLeakyReLU (slope 2−3) with quantized RPReLU (ablation) | ReActNet-18 · CIFAR-100 | Top-1 accuracy: 69.31 -> 69.35 | No clear effect |
| Set α to 2−2 rather than 2−3 (ablation) | ReActNet-18 · ImageNet | Top-1 accuracy: 60.55 -> 61.39 | Helped |
| Set α to 2−2 rather than 2−3 (ablation) | ReActNet-18 · CIFAR-10 | Top-1 accuracy: 91.69 -> 91.94 | Helped |
| Set α to 2−2 rather than 2−3 (ablation) | ReActNet-18 · CIFAR-100 | Top-1 accuracy: 68.35 -> 69.35 | Helped |
Surprising findings
- Quantized RPReLU improves ReActNet-A on ImageNet by 1.11% over the unquantized version, while having little effect on ReActNet-18.
- Simply reducing the first-layer stride of ReActNet-A on CIFAR gives 5.53% to 12.93% gains, showing architecture-specific sensitivity.
Limitations
- Accuracy drops about 1.1% on ReActNet-A ImageNet relative to BN-Free baseline.
- Mask layer must be present during training, adding training-time multiplications.
- Hardware evaluation is limited to synthesis of individual units on one FPGA, not full-network deployment.
Experimental setup
Datasets: ImageNet (ILSVRC12) CIFAR-10 CIFAR-100
Benchmarks: ImageNet CIFAR-10 CIFAR-100
Metrics: top-1 accuracy top-5 accuracy multiplication operand count LUT/slice/DSP hardware overhead
Baselines: XNOR-ResNet-18 Bi-ResNet-18 ReActNet-18 ReActNet-18 (BN-Free) ReActNet-A ReActNet-A (BN-Free) Bi-ResNet-34 BNN-ResNet-18 RLeakyReLU
Builds on: BN-Free BNN ReActNet Bi-Real Net XNOR-Net normalizer-free networks scaled weight standardization adaptive gradient clipping
Closest papers by meaning
- Binarized Neural Networks (NeurIPS 2016 · similarity 0.86)
- SA-BNN: State-Aware Binary Neural Network (AAAI 2021 · similarity 0.86)
- Binary Ensemble Neural Network: More Bits per Network or More Networks per Bit? (CVPR 2019 · similarity 0.85)
- BD-Net: Has Depth-Wise Convolution Ever Been Applied in Binary Neural Networks? (AAAI 2026 · similarity 0.85)
- Fast and Accurate Binary Neural Networks Based on Depth-Width Reshaping (AAAI 2023 · similarity 0.84)
- BiPer: Binary Neural Networks using a Periodic Function (CVPR 2024 · similarity 0.84)
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.