Research map / Architectures and efficiency / Quantization
Binary Ensemble Neural Network: More Bits per Network or More Networks per Bit?
Check 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 |
|---|---|---|---|
| Ensemble binary SB classifiers instead of using one classifier | SB on CIFAR-10 · CIFAR-10, NIN | Accuracy: 84.91% -> 89.12% | Helped |
| Add boosting rather than bagging, especially for compressed weak classifiers (ablation) | BENN on CIFAR-10 · CIFAR-10, NIN (Nano) | Accuracy: 57.12% -> 63.11%; up to 13.38% | Helped |
| Use warm-restart sequential training rather than independent training (ablation) | BENN on CIFAR-10 · CIFAR-10, NIN | Accuracy: up to 3.9% for bagging and 2.95% for boosting | Helped |
| Use warm-restart sequential training rather than independent training (ablation) | BENN on ImageNet · ImageNet | Accuracy | Hurt |
| Increase the number of ensemble classifiers | BENN on CIFAR-10 · CIFAR-10 | Accuracy: 4.21% to 24.16% over BNN; up to 1.52% over full-precision DNN | Helped |
| Increase the ensemble size from 5 to 32 (ablation) | BENN on CIFAR-10 · CIFAR-10 | Accuracy variation: variance by ~90% after 5 rounds and ~99% after 32 rounds | Helped |
| Select the best BNN checkpoint rather than the final training checkpoint (ablation) | Single BNN on CIFAR-10 · CIFAR-10 | Training accuracy oscillation: 2.94 -> 1.40 | Helped |
| Increase activation precision from binary to 2-bit (ablation) | BENN-AQB · CIFAR-10 | Accuracy | Helped |
| Keep the first and/or last layer full precision instead of binarizing all layers (ablation) | BENN · CIFAR-10 | Accuracy: loss reduces from 3.98% ~ 11.9% to 2.36% ~ 6.98% | Helped |
| Compress each base network to Tiny or Nano size (ablation) | BENN-SB · CIFAR-10 | Accuracy: compression reduces single BNN's accuracy by 7.71% (Tiny config) and 44.21% (Nano config) | Hurt |
| Increase ensemble size on ImageNet (ablation) | BENN-SB · ImageNet | Accuracy: 56%+ on AlexNet | Mixed |
| Replace bagging with boosting (ablation) | BENN-SB · ImageNet, AlexNet | Top-1 accuracy: 48.8% -> 50.2% (3 ensembles); 52.0% -> 54.3% (6 ensembles) | Helped |
Closest papers by meaning
- Binarized Neural Networks (NeurIPS 2016 · similarity 0.86)
- Compacting Binary Neural Networks by Sparse Kernel Selection (CVPR 2023 · similarity 0.86)
- Improving Accuracy of Binary Neural Networks Using Unbalanced Activation Distribution (CVPR 2021 · similarity 0.86)
- SA-BNN: State-Aware Binary Neural Network (AAAI 2021 · similarity 0.86)
- A&B BNN: Add&Bit-Operation-Only Hardware-Friendly Binary Neural Network (CVPR 2024 · similarity 0.85)
- Training Binary Neural Networks via Gaussian Variational Inference and Low-Rank Semidefinite Programming (NeurIPS 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.