Research map / Architectures and efficiency / Quantization
RTN: Reparameterized Ternary Network
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 |
|---|---|---|---|
| reparameterize ternary weights and activations with learnable scale and offset (ablation) | RTN on ImageNet ResNet-18 · ImageNet, ResNet-18 | top-1 accuracy: 62.4 -> 64.5 | Helped |
| reparameterize ternary activations with learnable scale and offset (ablation) | RTN on ImageNet AlexNet · ImageNet, AlexNet | top-1 accuracy: 52.6 -> 53.9 | Helped |
| reparameterize ternary activations with learnable scale and offset (ablation) | RTN on CIFAR-10 NIN · CIFAR-10, NIN | top-1 accuracy: 88.2 -> 89.6 | Helped |
| use activation scale factor only (ablation) | RTN on CIFAR-10 NIN · CIFAR-10, NIN | top-1 accuracy: 88.2 -> 88.5 | No clear effect |
| use activation offset factor only (ablation) | RTN on CIFAR-10 NIN · CIFAR-10, NIN | top-1 accuracy: 88.2 -> 89.1 | Helped |
| use ternary rather than quaternary vector multiplication | the vector multiplication circuit · ternary versus quaternary circuit | power consumption: 3.43$\times$ savings | Helped |
| use ternary rather than quaternary vector multiplication | the vector multiplication circuit · ternary versus quaternary circuit | area: 4.17$\times$ savings | Helped |
| use ternary rather than 32-bit floating-point vector multiplication | the vector multiplication circuit · ternary versus floating-point circuit | power consumption: 46.46$\times$ savings | Helped |
| use ternary rather than 32-bit floating-point vector multiplication | the vector multiplication circuit · ternary versus floating-point circuit | area: 89.17$\times$ savings | Helped |
Closest papers by meaning
- Trained Ternary Quantization (ICLR 2017 · similarity 0.90)
- TRQ: Ternary Neural Networks With Residual Quantization (AAAI 2021 · similarity 0.88)
- SYQ: Learning Symmetric Quantization for Efficient Deep Neural Networks (CVPR 2018 · similarity 0.88)
- Training Quantized Nets: A Deeper Understanding (NeurIPS 2017 · similarity 0.87)
- Incremental Network Quantization: Towards Lossless CNNs with Low-precision Weights (ICLR 2017 · similarity 0.87)
- HitNet: Hybrid Ternary Recurrent Neural Network (NeurIPS 2018 · similarity 0.87)
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.