Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Modular Meta-Learning with Shrinkage
Yutian Chen, Abram L. Friesen, Feryal Behbahani, Arnaud Doucet, David Budden, Matthew Hoffman, Nando de Freitas
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 learned shrinkage prior (σ-iMAML vs iMAML) | iMAML · Large-data augmented Omniglot | test accuracy (%): 72.8 -> 73.6 | Improved |
| add learned shrinkage prior (σ-Reptile vs Reptile) | Reptile · Large-data augmented Omniglot | test accuracy (%): 77.8 -> 78.9 | Improved |
| add shrinkage prior when few training instances per alphabet | iMAML and Reptile · Small-data augmented Omniglot | test accuracy | Improved |
| add shrinkage prior to Reptile with 4 mins speech data | Reptile · TTS naturalness, 4 mins | MOS: 1.93 -> 1.98 | Improved |
| add shrinkage prior to Reptile with 8 mins speech data | Reptile · TTS naturalness, 8 mins | MOS: 2.09 -> 2.28 | Improved |
| adapt only 12 of 123 modules selected by σ² instead of all modules (ablation) | Reptile · TTS voice similarity and naturalness | MOS: 2.09 -> 2.28 | Improved |
| adapt only the shrinkage-identified output layer rather than all layers (ablation) | σ-iMAML and σ-Reptile · Large-data augmented Omniglot | test accuracy | No clear effect |
| adapt penultimate layer only (ablation) | σ-MAML-style short-adaptation image classification · standard Omniglot / miniImageNet | test accuracy | Mixed |
| adapt only the first layer (ablation) | shrinkage method · Sinusoid regression | error | Improved |
| adapt all modules (SEA-All) instead of only speaker embedding (SEA-Emb) at 4 mins | SEA · TTS naturalness, 4 mins | MOS: 1.51 -> 1.41 | Hurt |
| adapt all modules (SEA-All) instead of only speaker embedding (SEA-Emb) at 8 mins | SEA · TTS naturalness, 8 mins | MOS: 1.57 -> 1.73 | Improved |
| meta-train with Reptile instead of finetuning SEA-All | WaveNet TTS · TTS naturalness, 8 mins | MOS: 1.73 -> 2.09 | Improved |
Closest papers by meaning
- Sample Efficient Adaptive Text-to-Speech (ICLR 2019 · similarity 0.87)
- Learn to Merge: Meta-Learning for Adaptive Multi-Task Model Merging (ICML 2026 · similarity 0.87)
- Adversarial Meta Sampling for Multilingual Low-Resource Speech Recognition (AAAI 2021 · similarity 0.87)
- Provable Meta-Learning with Low-Rank Adaptations (NeurIPS 2025 · similarity 0.86)
- MALIBO: Meta-learning for Likelihood-free Bayesian Optimization (ICML 2024 · similarity 0.85)
- Memory-Reduced Meta-Learning with Guaranteed Convergence (AAAI 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.
