Research map / Architectures and efficiency / CNN design
Projection Convolutional Neural Networks for 1-bit CNNs via Discrete Back Propagation
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 |
|---|---|---|---|
| Increase the number of projections from J=1 to J=4 (ablation) | PCNN with WRN-22, i=16 · CIFAR-10 | accuracy: 89.17 -> 92.79 | Helped |
| Increase the number of projections from J=1 to J=4 (ablation) | PCNN with WRN-22, i=16 · CIFAR-100 | accuracy: 62.66 -> 70.09 | Helped |
| Increase the number of projections from J=1 to J=4 (ablation) | PCNN with WRN-22, i=64 · CIFAR-10 | accuracy: 94.31 -> 95.39 | Helped |
| Increase the number of projections from J=1 to J=4 (ablation) | PCNN with WRN-22, i=64 · CIFAR-100 | accuracy: 76.93 -> 78.13 | Helped |
| Use two projections instead of one (ablation) | PCNN with ResNet18 and 32-bit activations · ImageNet | Top-1 accuracy: 63.5$^{\dag}$ -> 66.1$^{\ddag}$ | Helped |
| Use two projections instead of one (ablation) | PCNN with ResNet18 and 32-bit activations · ImageNet | Top-5 accuracy: 85.1$^{\dag}$ -> 86.7$^{\ddag}$ | Helped |
| Set lambda to 1e-4 rather than 1e-3, 1e-5, or 0 (ablation) | PCNN-22 with WRN-22 · CIFAR-10 | accuracy: 91.92, 92.79, 92.24, 91.52 | Helped |
| Set lambda to 1e-4 rather than 1e-3, 1e-5, or 0 (ablation) | PCNN-40 with WRN-40 · CIFAR-10 | accuracy: 92.85, 93.78, 93.65, 92.84 | Helped |
| Add projection loss with lambda=1e-4 instead of omitting it (ablation) | PCNN-22 · CIFAR-10 | accuracy: 91.52 -> 92.79 | Helped |
| Add projection loss with lambda=1e-4 instead of omitting it (ablation) | PCNN-40 · CIFAR-10 | accuracy: 92.84 -> 93.78 | Helped |
| Binarize activations as well as weights (ablation) | PCNN with ResNet18 and J=1 · ImageNet | Top-1 accuracy | Hurt |
| Binarize activations as well as weights (ablation) | PCNN with ResNet18 and J=1 · ImageNet | Top-5 accuracy | Hurt |
Closest papers by meaning
- Circulant Binary Convolutional Networks: Enhancing the Performance of 1-Bit DCNNs With Circulant Back Propagation (CVPR 2019 · similarity 0.86)
- BD-Net: Has Depth-Wise Convolution Ever Been Applied in Binary Neural Networks? (AAAI 2026 · similarity 0.85)
- Discrete Model Compression With Resource Constraint for Deep Neural Networks (CVPR 2020 · similarity 0.85)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · similarity 0.84)
- PICNN: A Pathway towards Interpretable Convolutional Neural Networks (AAAI 2024 · similarity 0.83)
- Network Memory Footprint Compression Through Jointly Learnable Codebooks and Mappings (ICLR 2024 · similarity 0.83)
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.