Research map / Continual learning / Online and replay-based
Piggyback: Adapting a Single Network to Multiple Tasks by Learning to Mask Weights
Arun Mallya, Dillon Davis, Svetlana Lazebnik
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 |
|---|---|---|---|
| learn binary piggyback masks instead of fine-tuning a separate network per task | VGG-16 · CUBS | top-1 error (%): 21.30 -> 20.99 | Improved |
| learn binary piggyback masks instead of fine-tuning a separate network per task | VGG-16 · Stanford Cars | top-1 error (%): 12.49 -> 11.87 | Improved |
| learn piggyback masks instead of PackNet | VGG-16 · CUBS | top-1 error (%): 22.30 -> 20.99 | Improved |
| delay task addition from first to last in PackNet | PackNet · CUBS | top-1 error (%): 22.30 -> 29.69 | Hurt |
| use piggyback masks on deeper backbone instead of individual networks | DenseNet-121 · WikiArt | top-1 error (%): 23.59 -> 29.56 | Hurt |
| initialize backbone from ImageNet instead of Places365 (ablation) | Piggyback ResNet-50 · CUBS | top-1 error (%): 28.50 -> 18.41 | Improved |
| initialize backbone from ImageNet instead of random weights (ablation) | Piggyback ResNet-50 · CUBS | top-1 error (%): 66.24 -> 18.41 | Improved |
| initialize backbone from ImageNet instead of random weights (ablation) | Piggyback ResNet-50 · Flowers | top-1 error (%): 71.17 -> 5.23 | Improved |
| train task-specific batchnorm parameters (ablation) | Piggyback ResNet-50 · WikiArt | top-1 error (%): 28.67 -> 25.92 | Improved |
| train task-specific batchnorm parameters (ablation) | Piggyback DenseNet-121 · WikiArt | top-1 error (%): 29.56 -> 25.90 | Improved |
| train task-specific batchnorm parameters (ablation) | Piggyback ResNet-50 · Sketch | top-1 error (%): 20.09 -> 19.82 | Improved |
| initialize mask weights with a constant and use Adam instead of other options (ablation) | Piggyback · fine-grained classification | accuracy: ~2% | Improved |
Later papers that compared with this method
3 later accepted papers report Piggyback as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Capacity-Agnostic Parameter Isolation for Continual Graph Learning (ICML 2026 · better on 0 of 1 shared result)
- Helpful or Harmful: Inter-Task Association in Continual Learning (ECCV 2022 · better on 3 of 3 shared results)
- KSM: Fast Multiple Task Adaption via Kernel-Wise Soft Mask Learning (CVPR 2021 · better on 0 of 1 shared result)
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
- PackNet: Adding Multiple Tasks to a Single Network by Iterative Pruning (CVPR 2018 · similarity 0.87)
- Piggyback GAN: Efficient Lifelong Learning for Image Conditioned Generation (ECCV 2020 · similarity 0.87)
- Adaptive Neural Networks for Efficient Inference (ICML 2017 · similarity 0.86)
- KSM: Fast Multiple Task Adaption via Kernel-Wise Soft Mask Learning (CVPR 2021 · similarity 0.86)
- Incremental Learning Through Deep Adaptation (TPAMI 2018 · similarity 0.85)
- RECAST: Reparameterized, Compact weight Adaptation for Sequential Tasks (ICLR 2025 · 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.
