Research map / Architectures and efficiency / CNN design
Fractional Skipping: Towards Finer-Grained Dynamic CNN Inference
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 dynamic fractional skipping instead of binary layer skipping | ResNet38 on CIFAR-10 at 20% computation · CIFAR-10, ResNet38, computation percentage 20% | accuracy: 89.26% -> 93.61% | Helped |
| Use dynamic fractional skipping instead of binary layer skipping | ResNet38 on CIFAR-100 at 20% computation · CIFAR-100, ResNet38, computation percentage 20% | accuracy: 60.38% -> 68.91% | Helped |
| Use dynamic fractional skipping instead of binary layer skipping | ResNet74 on CIFAR-100 at 20% computation · CIFAR-100, ResNet74, computation percentage 20% | accuracy: 65.09% -> 70.94% | Helped |
| Add dynamic fractional skipping to ResNet38 | ResNet38 on CIFAR-10 at 20% computation · CIFAR-10, ResNet38, computation percentage 20% | accuracy: 89.26% -> 93.61% | Helped |
| Dynamically change the sign of α rather than using a constant positive α (ablation) | DFS-ResNet74 · CIFAR-10, target computation percentage 40% | actual computation percentage: 8.20% -> 40.20% | Helped |
| Dynamically change the sign of α rather than using a constant positive α (ablation) | DFS-ResNet74 · CIFAR-10, target computation percentage 40% | accuracy: 93.12% -> 93.53% | Helped |
| Increase |α| from 1e-5 to 1e-4 (ablation) | DFS-ResNet74 · CIFAR-10, target computation percentage 40% | accuracy: 93.54% -> 93.31% | Hurt |
| Increase |α| from 1e-5 to 1e-3 (ablation) | DFS-ResNet74 · CIFAR-10, target computation percentage 50% | accuracy: 93.74% -> 93.42% | Hurt |
| Increase computation percentage from 4% to 6.25% | DFS-ResNet74 · CIFAR-10 | accuracy: 92.91% -> 93.54% | Helped |
| Use DFS rather than SkipNet at the same computation percentage | ResNet38 and ResNet74 on CIFAR-10 and CIFAR-100 · CIFAR-10 and CIFAR-100, computation percentage 20%-80% | accuracy: 20%-80%; slightly fluctuate within a range of 0.5% | Helped |
| Use DFS rather than static scalable quantization or HAQ | ResNet38 on CIFAR-10 · CIFAR-10, ResNet38 | accuracy: up to 1.2% over the scalable method and 0.2% over HAQ | Mixed |
| Use a larger α step size (ablation) | DFS-ResNet74 · CIFAR-10, target computation percentages 40% and 50% | actual computation percentage: 43.75% at target 40%; 53.40% at target 50% | Hurt |
Closest papers by meaning
- Skip-Convolutions for Efficient Video Processing (CVPR 2021 · similarity 0.87)
- Beyond Skip Connection: Pooling and Unpooling Design for Elimination Singularities (AAAI 2025 · similarity 0.86)
- Reducing Fine-Tuning Memory Overhead by Approximate and Memory-Sharing Backpropagation (ICML 2024 · similarity 0.85)
- Adaptive Depth Networks with Skippable Sub-Paths (NeurIPS 2024 · similarity 0.85)
- Towards Cheaper Inference in Deep Networks with Lower Bit-Width Accumulators (ICLR 2024 · similarity 0.85)
- Quantization without Tears (CVPR 2025 · similarity 0.85)
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.