Research map / Video / Action recognition
Stochastic Backpropagation: A Memory Efficient Strategy for Training Video Models
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 |
|---|---|---|---|
| Apply SBP with keep-ratio 0.25 instead of full E2E training | Swin-T · Kinetics-400 | Top-1 accuracy (%): 78.8 -> 78.2 | Hurt |
| Apply SBP with keep-ratio 0.25 instead of full E2E training | Swin-T · Epic-Kitchens-55 | Top-1 accuracy (%): 46.0 -> 45.2 | Hurt |
| Preserve all forward paths and drop selected backward paths, rather than dropping both paths | Swin-T · Epic-Kitchens-55 | Top-1 accuracy (%): 38.9 -> 45.2 | Helped |
| Apply SBP with keep-ratio 0.25 instead of full E2E training | Swin-B · Kinetics-400 | Top-1 accuracy (%): 82.7 -> 81.9 | Hurt |
| Apply SBP with keep-ratio 0.25 instead of full E2E training | Swin-B · Epic-Kitchens-55 | Top-1 accuracy (%): 49.4 -> 47.3 | Hurt |
| Apply SBP with keep-ratio 0.25 instead of full E2E training | G-TAD · THUMOS'14 temporal action localization | Average mAP (%): 31.17 -> 30.14 | Hurt |
| Apply SBP with keep-ratio 0.125 instead of full E2E training | G-TAD · THUMOS'14 temporal action localization | Average mAP (%): 31.17 -> 29.41 | Hurt |
| Use 3D-checkerboard spatial-temporal sampling instead of temporal-only sampling (ablation) | Swin-T SBP · Kinetics-400 | Top-1 accuracy (%): 78.2 -> 78.3 | No clear effect |
| Use 3D-checkerboard spatial-temporal sampling instead of temporal-only sampling (ablation) | Swin-T SBP · Epic-Kitchens-55 | Top-1 accuracy (%): 45.2 -> 45.4 | No clear effect |
| Use a diverse-feature sampler rather than the uniform random sampler (ablation) | SBP for online action detection · THUMOS'14 | mAP | No clear effect |
| Use a diverse-gradient sampler rather than the uniform random sampler (ablation) | SBP for online action detection · THUMOS'14 | mAP | No clear effect |
| Set spatial-model learning rate about 10× below the temporal-model learning rate (ablation) | SBP spatial-then-temporal models · THUMOS'14 online action detection | mAP | Helped |
Closest papers by meaning
- An In-depth Study of Stochastic Backpropagation (NeurIPS 2022 · similarity 0.93)
- Reducing Fine-Tuning Memory Overhead by Approximate and Memory-Sharing Backpropagation (ICML 2024 · similarity 0.88)
- Sideways: Depth-Parallel Training of Video Models (CVPR 2020 · similarity 0.87)
- Backprop with Approximate Activations for Memory-efficient Network Training (NeurIPS 2019 · similarity 0.86)
- Accelerated training through iterative gradient propagation along the residual path (ICLR 2025 · similarity 0.85)
- MEET: Towards Memory-Efficient Temporal Sparse Deep Neural Networks (CVPR 2025 · 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.