Research map / Architectures and efficiency / Pruning and efficient tuning
AMC: AutoML for Model Compression and Acceleration on Mobile Devices
Yihui He, Ji Lin, Zhijian Liu, Hanrui Wang, Li-Jia Li, Song Han
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 |
|---|---|---|---|
| learned RL compression policy instead of uniform handcrafted ratios | channel pruning, ResNet-56 · CIFAR-10 (50% FLOPs) | accuracy after fine-tuning (%): 89.8 -> 91.9 | Improved |
| learned RL policy instead of best handcrafted policy | channel pruning, Plain-20 · CIFAR-10 (50% FLOPs) | accuracy before fine-tuning (%): 83.9 -> 86.0 | Improved |
| learned RL policy instead of CP expert-tuned ratios | VGG-16 channel pruning · ImageNet (20% FLOPs) | top-1 accuracy change (%): -1.7 -> -1.4 | Improved |
| learned RL policy instead of uniform width multiplier | MobileNet · ImageNet (50% FLOPs) | top-1 accuracy change (%): -2.5 -> -0.4 | Improved |
| learned RL policy instead of uniform width multiplier | MobileNet-V2 · ImageNet (70% FLOPs) | top-1 accuracy change (%): -2.0 -> -1.0 | Improved |
| use latency-constrained search instead of FLOPs-constrained | AMC on MobileNet · Pixel 1 Android latency | speedup: 1.81x -> 1.95x | Improved |
| use latency-constrained search instead of FLOPs-constrained | AMC on MobileNet · Titan XP GPU | speedup: 1.43x -> 1.53x | Improved |
| AMC pruning instead of handcrafted pruning | MobileNet · ImageNet val | top-1 accuracy (%): 67.2 -> 70.5 | Improved |
| AMC policy instead of NetAdapt heuristic | MobileNet · Pixel 1 latency | latency (ms): 70.0ms -> 63.3ms | Improved |
| iterative fine-grained pruning with AMC instead of expert ratios | ResNet-50 · ImageNet val | compression ratio: 3.4x -> 5x | Improved |
| AMC compression of VGG-16 backbone instead of handcrafted 4x pruning | Faster R-CNN VGG16 · PASCAL VOC 2007 | mAP [0.5, 0.95] (%): 36.5 -> 37.2 | Improved |
| AMC compression instead of no compression | Faster R-CNN VGG16 · PASCAL VOC 2007 | mAP (%): 68.7 -> 68.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- NetAdapt — NetAdapt: Platform-Aware Neural Network Adaptation for Mobile Applications (ECCV 2018) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Auto Graph Encoder-Decoder for Neural Network Pruning (ICCV 2021 · similarity 0.87)
- Learning Accurate Performance Predictors for Ultrafast Automated Model Compression (IJCV 2023 · similarity 0.86)
- Auto-Compressing Networks (NeurIPS 2025 · similarity 0.84)
- PALMBENCH: A COMPREHENSIVE BENCHMARK OF COMPRESSED LARGE LANGUAGE MODELS ON MOBILE PLATFORMS (ICLR 2025 · similarity 0.83)
- EvoPress: Accurate Dynamic Model Compression via Evolutionary Search (ICML 2025 · similarity 0.83)
- Fix the Structural Bottleneck: Context Compression via Explicit Information Transmission (NeurIPS 2026 · similarity 0.83)
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.
