Research map / Architectures and efficiency / RNNs and normalization
SmartFRZ: An Efficient Training Framework using Attention-Based Layer Freezing
Sheng Li, Geng Yuan, Yue Dai, Youtao Zhang, Yanzhi Wang, Xulong Tang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| attention-based layer freezing (SmartFRZ) instead of full training | ResNet50 fine-tuning · CIFAR-100 fine-tuning | computation (TFLOPs) and accuracy: 12,360 -> 6,398 TFLOPs; accuracy 81.68% -> 81.73% | Improved |
| attention-based layer freezing instead of linear freezing | VGG11 fine-tuning · CIFAR-100 fine-tuning | accuracy: 73.26% -> 74.66% (+1.4) | Improved |
| attention-based layer freezing instead of AutoFreeze | VGG11 fine-tuning · CIFAR-10 fine-tuning | computation (TFLOPs): 13,350 -> 10,059 TFLOPs | Improved |
| attention-based layer freezing instead of full training | ResNet50 training from scratch · CIFAR-100 from scratch | accuracy and time: 711 -> 538 min; accuracy 77.48% -> 77.64% | Improved |
| attention-based layer freezing instead of full training | ResNet50 training from scratch · ImageNet from scratch | accuracy and time: 76.89% -> 76.80% accuracy (-0.09); 780 -> 621 min | Mixed |
| attention-based layer freezing instead of AutoFreeze and linear freezing | ResNet50 training from scratch · CIFAR-100 from scratch | accuracy: 72.82% (AutoFreeze), 74.99% (linear) -> 77.64% | Improved |
| attention-based layer freezing instead of full training | DeiT-T fine-tuning · CIFAR-100 fine-tuning | accuracy: 85.03% -> 84.82% (-0.21); time 14,628 -> 9,599 s | Hurt |
| attention-based layer freezing instead of full training | VGG11 fine-tuning · CIFAR-100 fine-tuning | accuracy: 74.95% -> 74.66% (-0.29); time 2,703 -> 1,831 s | Hurt |
| attention-based layer freezing instead of full training | BERT-base fine-tuning · MRPC and CoLA fine-tuning | accuracy and time: MRPC 86.51% -> 86.76%, 316 -> 215 s; CoLA 56.65% -> 56.66%, 723 -> 423 s | Improved |
| attention window size varied from 20 to 50 (ablation) | SmartFRZ predictor (ResNet50) · CIFAR-100 fine-tuning | accuracy: 81.17% (20) -> 81.73% (30) -> 82.11% (50) | Improved |
| tailored layer size reduced from 1024 to 512 (ablation) | SmartFRZ predictor · CIFAR-100 fine-tuning (ResNet50, VGG11) | accuracy: ResNet50 81.73% -> 81.28%; VGG11 74.66% -> 73.98% | Hurt |
| AutoFreeze gradient-norm overhead removed (SmartFRZ uses predictor) | AutoFreeze baseline · fine-tuning time | time share | No clear effect |
Closest papers by meaning
- On The Concurrence of Layer-wise Preconditioning Methods and Provable Feature Learning (ICML 2025 · similarity 0.82)
- To Cool or not to Cool? Temperature Network Meets Large Foundation Models via DRO (ICML 2024 · similarity 0.82)
- DASH: Warm-Starting Neural Network Training in Stationary Settings without Loss of Plasticity (NeurIPS 2024 · similarity 0.81)
- Layerwise Progressive Freezing: A Training Scaffold for Depth-Scalable Binary Networks (NeurIPS 2026 · similarity 0.81)
- Efficient stagewise pretraining via progressive subnetworks (ICLR 2025 · similarity 0.81)
- Frozen Feature Augmentation for Few-Shot Image Classification (CVPR 2024 · similarity 0.81)
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.
