Research map / Architectures and efficiency / CNN design
ReSprop: Reuse Sparsified Backpropagation
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 |
|---|---|---|---|
| Add a warm-up phase that gradually increases reuse-sparsity to 90% (ablation) | ReSprop · CIFAR-100, ResNet34 | validation accuracy: 69.12 ± 0.13 -> 75.14 ± 0.17 | Helped |
| Add a warm-up phase that gradually increases reuse-sparsity to 90% (ablation) | ReSprop · CIFAR-10, ResNet34 | validation accuracy: 91.61 ± 0.09 -> 94.36 ± 0.07 | Helped |
| Add a warm-up phase that gradually increases reuse-sparsity to 90% (ablation) | ReSprop · ImageNet, ResNet34 | top-1 validation accuracy: 63.78 -> 72.44 | Helped |
| Use 90% gradient reuse-sparsity with a warm-up phase | W-ReSprop · ImageNet, ResNet34 | top-1 validation accuracy: 73.34 -> 72.44 | Hurt |
| Use 90% gradient reuse-sparsity with a warm-up phase | W-ReSprop · ImageNet, WRN-50-2 | top-1 validation accuracy: 78.88 -> 77.93 | Hurt |
| Use 90% gradient reuse-sparsity with a warm-up phase | W-ReSprop · ImageNet, VGG16 | top-1 validation accuracy: 70.50 -> 69.46 | Hurt |
| Run ReSprop on GSCN rather than standard training on GSCN | GSCN · ImageNet backward pass, ResNet34 | speedup: 1.81 -> 8.01 | Helped |
| Run ReSprop on GSCN rather than standard training on GSCN | GSCN · ImageNet backward pass, VGG16 | speedup: 1.27 -> 7.21 | Helped |
| Use stochastic gradient sampling instead of storing full mini-batch gradients (ablation) | ReSprop, ResNet-18 · CIFAR-10, 90% reuse-sparsity | validation accuracy: 91.10 ± 0.11 -> 91.43 ± 0.11 | No clear effect |
| Use average previous mini-batch gradients instead of storing full mini-batch gradients (ablation) | ReSprop, ResNet-18 · CIFAR-10, 90% reuse-sparsity | validation accuracy: 91.10 ± 0.11 -> 91.82 ± 0.07 | No clear effect |
| Replace gradient reuse with setting low-magnitude gradients to zero | MeProp, ResNet-18 · CIFAR-10 | validation accuracy: 84.21 ± 0.09 -> 64.29 ± 0.07 | Hurt |
| Increase reuse-sparsity from 50% to 90% (ablation) | W-ReSprop, ResNet34 · CIFAR-100 | validation accuracy: 76.4 ± 0.11 -> 75.14 ± 0.17 | Hurt |
Closest papers by meaning
- SparseProp: Efficient Sparse Backpropagation for Faster Training of Neural Networks at the Edge (ICML 2023 · similarity 0.91)
- meProp: Sparsified Back Propagation for Accelerated Deep Learning with Reduced Overfitting (ICML 2017 · similarity 0.88)
- Backpropagation-Free Deep Learning with Recursive Local Representation Alignment (AAAI 2023 · similarity 0.88)
- Accelerating Sparse Convolution with Column Vector-Wise Sparsity (NeurIPS 2022 · similarity 0.87)
- Accelerated training through iterative gradient propagation along the residual path (ICLR 2025 · similarity 0.86)
- Preserving Deep Representations in One-Shot Pruning: A Hessian-Free Second-Order Optimization Framework (ICLR 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.