Research map / Architectures and efficiency / RNNs and normalization
YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information
Chien-Yao Wang, I-Hau Yeh, Hong-Yuan Mark Liao
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 |
|---|---|---|---|
| add PGI (auxiliary reversible branch) to GELAN-E (ablation) | GELAN-E · MS COCO val | AP50:95 (%): 55.0% -> 55.6% | Improved |
| add PGI to GELAN-C (ablation) | GELAN-C · MS COCO val | AP50:95 (%): 52.5% -> 53.0% | Improved |
| add PGI to GELAN-S (ablation) | GELAN-S · MS COCO val | AP50:95 (%): 46.7% -> 46.8% | Improved |
| add deep supervision to lightweight GELAN-S (ablation) | GELAN-S · MS COCO val | AP50:95 (%): 46.7% -> 46.5% | Hurt |
| add deep supervision to GELAN-C (ablation) | GELAN-C · MS COCO val | AP50:95 (%): 52.5% -> 52.5% | No clear effect |
| add deep supervision to deep GELAN-E (ablation) | GELAN-E · MS COCO val | AP50:95 (%): 55.0% -> 55.3% | Improved |
| use CSP blocks instead of conv blocks in GELAN-S (ablation) | GELAN-S · MS COCO val | AP50:95 (%): 44.8% -> 45.5% | Improved |
| replace conv blocks with Res blocks in GELAN-S (ablation) | GELAN-S · MS COCO val | AP50:95 (%): 44.8% -> 44.3% | Hurt |
| increase ELAN depth from 1 to 2 in GELAN-C (ablation) | GELAN-C · MS COCO val | AP50:95 (%): 50.7% -> 52.5% | Improved |
| increase ELAN or CSP depth beyond 2 (ablation) | GELAN-C · MS COCO val | AP50:95 (%): 52.5% -> 53.3% | No clear effect |
| add lead-head guided assignment to PGI auxiliary supervision (ablation) | GELAN-C · MS COCO val | AP50:95 (%): 52.9% -> 53.0% | Improved |
| add PFH (deep supervision) on backbone of GELAN-C (ablation) | GELAN-C · MS COCO val | AP50:95 (%): 52.5% -> 52.5% | No clear effect |
Closest papers by meaning
- Gradients as Features for Deep Representation Learning (ICLR 2020 · similarity 0.83)
- Learned Optimizers that Scale and Generalize (ICML 2017 · similarity 0.83)
- Disentangling Trainability and Generalization in Deep Neural Networks (ICML 2020 · similarity 0.83)
- Fast and Slow Gradient Approximation for Binary Neural Network Optimization (AAAI 2025 · similarity 0.82)
- Information Bottleneck Analysis of Deep Neural Networks via Lossy Compression (ICLR 2024 · similarity 0.82)
- GIO: Gradient Information Optimization for Training Dataset Selection (ICLR 2024 · similarity 0.82)
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.
