Research map / Architectures and efficiency / Pruning and efficient tuning
Back Razor: Memory-Efficient Transfer Learning by Self-Sparsified Backpropagation
Ziyu Jiang, Xuxi Chen, Xueqin Huang, Xianzhi Du, Denny Zhou, Zhangyang Wang
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 |
|---|---|---|---|
| prune activations only in backward instead of both forward and backward (ablation) | Back Razor · Aircraft / Pets (CNN) | top-1 accuracy: accuracy below 5% at 70% pruning for symmetric pruning vs high accuracy at 99% for backward-only | Improved |
| raise sparsity from 90% to 99% (ablation) | Back Razor · Pets | top-1 accuracy: 92.9 -> 92.5 | No clear effect |
| raise sparsity from 90% to 99% (ablation) | Back Razor · Aircraft | top-1 accuracy: 87.6 -> 80.7 | Hurt |
| raise sparsity from 90% to 97% (ablation) | Back Razor · Aircraft | top-1 accuracy: 87.6 -> 85.2 | Hurt |
| raise sparsity from 90% to 99% (memory) (ablation) | Back Razor · CIFAR100 CNN | training memory: 42MB -> 39MB | No clear effect |
| replace TinyTL@320 with Back Razor@90% | ProxylessNAS-Mobile · Aircraft | top-1 accuracy: 82.3 -> 87.6 | Improved |
| replace FT-Norm+Last with Back Razor@90% | ProxylessNAS-Mobile · Aircraft | top-1 accuracy: 68.1 -> 87.6 | Improved |
| replace SWAT@80% with Back Razor@90% | ResNet-18 · CIFAR100 | top-1 accuracy: 79.6 -> 82.2 | Improved |
| replace SWAT@80% with Back Razor@90% | ResNet-18 · CIFAR10 | top-1 accuracy: 95.8 -> 96.4 | Improved |
| raise ViT sparsity from 80% to 95% (ablation) | Back Razor · Aircraft (ViT-B/16) | top-1 accuracy: 77.3 -> 74.3 | Hurt |
| raise ViT sparsity from 80% to 95% (ablation) | Back Razor · CIFAR10 (ViT-B/16) | top-1 accuracy: 98.9 -> 98.8 | No clear effect |
| combine Back Razor@80% with Mesa quantization (ablation) | Back Razor · CIFAR100 (ViT-B/16) | training memory: 4565MB -> 2937MB | Improved |
Closest papers by meaning
- Reducing Fine-Tuning Memory Overhead by Approximate and Memory-Sharing Backpropagation (ICML 2024 · similarity 0.88)
- Backprop with Approximate Activations for Memory-efficient Network Training (NeurIPS 2019 · similarity 0.87)
- Accelerated training through iterative gradient propagation along the residual path (ICLR 2025 · similarity 0.84)
- A Closer Look at the Few-Shot Adaptation of Large Vision-Language Models (CVPR 2024 · similarity 0.84)
- Forward Learning with Top-Down Feedback: Empirical and Analytical Characterization (ICLR 2024 · similarity 0.84)
- Knowledge Transfer from Vision Foundation Models for Efficient Training of Small Task-specific Models (ICML 2024 · 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.
