Research map / Continual learning / Continual fine-tuning of LLMs
Task-Distributionally Robust Data-Free Meta-Learning
Zixuan Hu, Yongxian Wei, Li Shen, Zhenyi Wang, Baoyuan Wu, Chun Yuan, Dacheng Tao
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 memory buffer to vanilla synthetic-task training (ablation) | TDFML · CIFAR-FS 5-way 1-shot | LAST accuracy: 26.76 -> 28.33 | No clear effect |
| add task interpolation to memory buffer (ablation) | TDFML · CIFAR-FS 5-way 1-shot | LAST accuracy: 28.33 -> 38.42 | Improved |
| add soft-label supervision from pre-trained models (ablation) | TDFML · CIFAR-FS 5-way 1-shot | BEST accuracy: 38.76 -> 40.80 | Improved |
| add automatic model selection (deceptive-label models, 40% pollution) (ablation) | TDFML · CIFAR-FS 5-way 5-shot | accuracy: 47.03 -> 49.72 | Improved |
| add automatic model selection (deceptive-accuracy models, 80% pollution) (ablation) | TDFML · CIFAR-FS 5-way 5-shot | accuracy: 42.95 -> 45.85 | Improved |
| use task mixup hybrid strategy (+mixup) (ablation) | TDFML-ProtoNet · CIFAR-FS 5-way 1-shot | BEST accuracy: 40.80 -> 41.91 | Improved |
| use task mixup hybrid strategy (+mixup) (ablation) | TDFML-ProtoNet · CUB 5-way 5-shot | LAST accuracy: 47.64 -> 48.78 | Improved |
| use ProtoNet instead of ANIL as meta-learner (ablation) | TDFML · CIFAR-FS 1-shot vs MiniImageNet 5-shot | BEST accuracy: CIFAR-FS 1-shot 40.39 -> 40.80; MiniImageNet 5-shot 43.63 -> 42.93 | Mixed |
| replace PURER-ProtoNet with TDFML-ProtoNet | meta-learner · CIFAR-FS 5-way 1-shot | LAST accuracy: 27.01 -> 40.28 | Improved |
| increase pollution rate from 10% to 80% (deceptive label) | PURER · CIFAR-FS 5-way 5-shot | accuracy: 48.81 -> 41.02 | Hurt |
| compare with data-free KD method MAD | TDFML · CIFAR-FS 5-way 5-shot | BEST accuracy: 35.19 -> 58.25 | Improved |
| compare with continual meta-learning SDML | TDFML · CIFAR-FS 5-way 5-shot | BEST accuracy: 49.25 -> 58.25 | Improved |
Closest papers by meaning
- Task-Robust Model-Agnostic Meta-Learning (NeurIPS 2020 · similarity 0.89)
- A Simple Yet Effective Strategy to Robustify the Meta Learning Paradigm (NeurIPS 2023 · similarity 0.88)
- Task Agnostic Meta-Learning for Few-Shot Learning (CVPR 2019 · similarity 0.87)
- Task Groupings Regularization: Data-Free Meta-Learning with Heterogeneous Pre-trained Models (ICML 2024 · similarity 0.87)
- Theoretical Investigations and Practical Enhancements on Tail Task Risk Minimization in Meta Learning (NeurIPS 2024 · similarity 0.86)
- MTL-UE: Learning to Learn Nothing for Multi-Task Learning (ICML 2025 · similarity 0.86)
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.
