Research map / Architectures and efficiency / Pruning and efficient tuning
Fire Together Wire Together: A Dynamic Pruning Approach With Self-Supervised Mask Prediction
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 |
|---|---|---|---|
| Use self-supervised dynamic filter routing instead of static inference | VGG16-BN · CIFAR-10, 50% FLOPs reduction | Top-1 accuracy (%): 92.00 -> 93.80 | Helped |
| Decouple prediction-head training from backbone gradients (ablation) | FTWT ResNet34 · ImageNet | Top-1 accuracy (%): 70.06 -> 71.71 | Helped |
| Add softmax normalization to decision heads (ablation) | FTWT VGG16-BN · CIFAR-10, Gaussian blur, first reported setting | Brier score: 0.19 -> 0.12 | Helped |
| Add softmax normalization to decision heads (ablation) | FTWT VGG16-BN · CIFAR-10, Gaussian blur, second reported setting | Brier score: 0.48 -> 0.18 | Helped |
| Add softmax normalization to decision heads (ablation) | FTWT VGG16-BN · CIFAR-10, additive noise, first reported setting | Brier score: 0.12 -> 0.12 | No clear effect |
| Add softmax normalization to decision heads (ablation) | FTWT VGG16-BN · CIFAR-10, additive noise, second reported setting | Brier score: 0.13 -> 0.12 | Helped |
| Allow input-dependent filter routes rather than fixed static routes (ablation) | VGG16-BN · CIFAR-10, 85% FLOPs reduction | Top-1 accuracy (%): 91.12 -> 92.75 | Helped |
| Allow input-dependent filter routes rather than fixed static routes (ablation) | ResNet56 · CIFAR-10, 70% FLOPs reduction | Top-1 accuracy (%): 91.61 -> 92.09 | Helped |
| Use joint rather than decoupled training (ablation) | FTWT on CIFAR · CIFAR-10, lower mass thresholds / high pruning | Top-1 accuracy (%) | Hurt |
| Use dynamic pruning rather than static pruning at similar computation | FTWT · ImageNet ResNet34, approximately 25 FLOPs reduction | Top-1 accuracy drop (%): 0.05 | Helped |
Closest papers by meaning
- Accelerate CNNs from Three Dimensions: A Comprehensive Pruning Framework (ICML 2021 · similarity 0.89)
- Dynamic Structure Pruning for Compressing CNNs (AAAI 2023 · similarity 0.89)
- DPFPS: Dynamic and Progressive Filter Pruning for Compressing Convolutional Neural Networks from Scratch (AAAI 2021 · similarity 0.89)
- Structured Pruning of Neural Networks With Budget-Aware Regularization (CVPR 2019 · similarity 0.89)
- Dynamic Sparse Training: Find Efficient Sparse Network From Scratch With Trainable Masked Layers (ICLR 2020 · similarity 0.89)
- Rethinking the Smaller-Norm-Less-Informative Assumption in Channel Pruning of Convolution Layers (ICLR 2018 · similarity 0.89)
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.