Research map / Architectures and efficiency / Pruning and efficient tuning
Zoo-Tuning: Adaptive Transfer from A Zoo of Models
Yang Shu, Zhi Kou, Zhangjie Cao, Jianmin Wang, Mingsheng Long
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 |
|---|---|---|---|
| adaptive aggregation (Zoo-Tuning) of source models vs single-model fine-tuning | ResNet-50 model zoo · 7 classification tasks | Avg. Acc.: 85.16 vs 83.41 best single (ImageNet sup.) | Improved |
| temporal-ensemble unified gating (lite Zoo-Tuning) instead of per-sample gating | Zoo-Tuning · 7 classification tasks | Avg. Acc.: 84.51 vs 85.16 | Hurt |
| removing channel alignment layer T (ablation) | Zoo-Tuning · COCO-70 | accuracy: 83.92 vs 84.91 | Hurt |
| removing channel alignment layer T (ablation) | Zoo-Tuning · WFLW | NME: 4.64 vs 4.58 | Hurt |
| learned adaptive gating vs average aggregation (ablation) | Zoo-Tuning · COCO-70 | accuracy: 84.91 vs 80.53 | Improved |
| learned adaptive gating vs average aggregation (ablation) | Zoo-Tuning · WFLW | NME: 4.58 vs 4.75 | Improved |
| 5 ImageNet supervised models aggregated instead of 5 diverse source models (same parameter count) (ablation) | Zoo-Tuning · Aircraft / COCO-70 | accuracy: 84.92 vs 86.54 (Aircraft); 82.13 vs 84.91 (COCO) | Hurt |
| increasing number of source models from 2 to 5 (ablation) | Zoo-Tuning · CIFAR-100 | accuracy: 78.62 (2 models) to 83.77 (5 models) | Improved |
| fine-tuning each source model separately (no aggregation) on RL target tasks | single-model transfer in Atari RL · Atari transfer (Alien, Gopher, JamesBond) | score | No clear effect |
| Zoo-Tuning vs Knowledge Flow on facial landmarks | multi-model transfer · 300W/WFLW/COFW | NME: 4.58 vs 5.28 on WFLW | Improved |
| Zoo-Tuning vs ensemble on facial landmarks | multi-model transfer · 300W/COFW | NME: 3.41 vs 3.33 (300W); 3.51 vs 3.46 (COFW) | Hurt |
Later papers that compared with this method
1 later accepted paper reports Zoo Tuning as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Deep Model Reassembly (NeurIPS 2022 · better on 3 of 3 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
- Model Zoos: A Dataset of Diverse Populations of Neural Network Models (NeurIPS 2022 · similarity 0.86)
- Tuning Large Neural Networks via Zero-Shot Hyperparameter Transfer (NeurIPS 2021 · similarity 0.85)
- NASOA: Towards Faster Task-Oriented Online Fine-Tuning With a Zoo of Models (ICCV 2021 · similarity 0.85)
- Growing a Brain: Fine-Tuning by Increasing Model Capacity (CVPR 2017 · similarity 0.85)
- Fine-Tune Once, Reuse Across Models: Bayesian Task-Update Factors and Approximations (ICML 2026 · similarity 0.85)
- UniPT: Universal Parallel Tuning for Transfer Learning with Efficient Parameter and Memory (CVPR 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.
