Research map / Architectures and efficiency / CNN design
BlockDrop: Dynamic Inference Paths in Residual Networks
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 learned instance-specific block dropping instead of a static heuristic | ResNet-110 · CIFAR-10 | accuracy: 93.6 vs 90.1 | Helped |
| Jointly finetune the policy and ResNet after curriculum training (ablation) | BlockDrop ResNet-110 · CIFAR-10 | accuracy: 75.4 -> 93.6 | Helped |
| Jointly finetune the policy and ResNet after curriculum training (ablation) | BlockDrop ResNet-110 · CIFAR-100 | accuracy: 72.1 -> 73.7 | Helped |
| Increase γ from 2 to 10, favoring accuracy over block savings (ablation) | BlockDrop-110 · CIFAR-10 | accuracy: 91.9 -> 93.6 | Helped |
| Increase γ from 2 to 10, favoring accuracy over block savings (ablation) | BlockDrop · ImageNet validation | accuracy (top-1): 75.2 -> 76.8 | Helped |
| Use γ=5 rather than the full ResNet-101 inference path | BlockDrop · ImageNet validation | FLOPs: 1.56E+10 -> 1.25E+10 | Helped |
| Make block decisions sequentially instead of predicting the entire path at once (ablation) | BlockDrop · CIFAR-10 | inference time: 10.9 -> 29.1 | Hurt |
| Use instance-specific learned paths rather than fixed block subsets | ResNet-110 · CIFAR-100 | accuracy: 73.7 vs 69.9 | Helped |
| Use a larger accuracy-penalty setting, γ=10 instead of γ=5 (ablation) | BlockDrop · ImageNet validation | FLOPs: 1.25E+10 -> 1.47E+10 | Hurt |
Closest papers by meaning
- DropBlock: A regularization method for convolutional networks (NeurIPS 2018 · similarity 0.88)
- Blockout: Dynamic Model Selection for Hierarchical Deep Networks (CVPR 2016 · similarity 0.88)
- ResNEsts and DenseNEsts: Block-based DNN Models with Improved Representation Guarantees (NeurIPS 2021 · similarity 0.88)
- Adaptive Neural Networks for Efficient Inference (ICML 2017 · similarity 0.86)
- Residual Alignment: Uncovering the Mechanisms of Residual Networks (NeurIPS 2023 · similarity 0.86)
- Depthwise Hyperparameter Transfer in Residual Networks: Dynamics and Scaling Limit (ICLR 2024 · 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.