Research map / Architectures and efficiency / Pruning and efficient tuning
OTOv2: Automatic, Generic, User-Friendly
Read the paperCheck 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 |
|---|---|---|---|
| Replace HSPG with DHSPG as the structured-sparsity optimizer (CARN x2, 50% target group sparsity) (ablation) | OTOv2 on CARN · Set14 PSNR | PSNR: 33.2 vs 33.0 (HSPG) | Helped |
| Replace HSPG with DHSPG as the structured-sparsity optimizer (CARN x2) (ablation) | OTOv2 on CARN · B100 and Urban100 PSNR | PSNR: B100 31.9 vs 31.6; Urban100 31.1 vs 30.9 | Helped |
| Replace HSPG with DHSPG in OTOv1 framework (BERT on SQuAD) (ablation) | BERT · SQuAD F1 | F1-score: 83.8-87.7% vs 82.0-84.1% for HSPG/ProxSSI | Helped |
| Automatic ZIG partition + DHSPG one-shot training and compression | DemoNet · Fashion-MNIST | Top-1 accuracy: -0.2 points (84.5 to 84.3) at 24.0% FLOPs | No clear effect |
| Automatic ZIG partition + DHSPG one-shot training and compression | ConvNeXt-Tiny · ImageNet | Top-1 accuracy: -0.9 points (82.0 to 81.1) at 52.8% FLOPs | Hurt |
| Automatic ZIG partition + DHSPG one-shot training and compression | DenseNet121 · CIFAR100 | Top-1 accuracy: -1.5 points (77.0 to 75.5) at 20.8% FLOPs | Hurt |
| DHSPG one-shot training with ZIG pruning on VGG16 (85% group sparsity) | VGG16 on CIFAR10 · CIFAR10 | FLOPs reduction to 13.4% with equal accuracy: 91.0% accuracy at 13.4% FLOPs vs OTOv1 91.0% at 16.3% FLOPs | Helped |
| Aggressive group sparsity (90%) on ResNet50 | ResNet50 on CIFAR10 · CIFAR10 | Top-1 accuracy: 93.0% at 2.2% FLOPs vs 94.4% for OTOv1 at 12.8% FLOPs | Hurt |
Closest papers by meaning
- Only Train Once: A One-Shot Neural Network Training And Pruning Framework (NeurIPS 2021 · similarity 0.89)
- On Compressing Deep Models by Low Rank and Sparse Decomposition (CVPR 2017 · similarity 0.87)
- MCNC: Manifold-Constrained Reparameterization for Neural Compression (ICLR 2025 · similarity 0.86)
- Effective Interplay between Sparsity and Quantization: From Theory to Practice (ICLR 2025 · similarity 0.86)
- Auto-Train-Once: Controller Network Guided Automatic Network Pruning from Scratch (CVPR 2024 · similarity 0.86)
- Language Models as Zero-shot Lossless Gradient Compressors: Towards General Neural Parameter Prior Models (NeurIPS 2024 · 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.