Research map / Probabilistic and scientific modeling / Variational and generative sampling
Empirical Bayes Transductive Meta-Learning with Synthetic Gradients
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 synthetic-gradient transductive steps (K=3) instead of using the pretrained feature only (K=0) (ablation) | SIB with Conv-4-64 · MiniImageNet 5-way 1-shot | accuracy: 50.0$\pm$0.4\% -> 58.0$\pm$0.6\% | Helped |
| Add synthetic-gradient transductive steps (K=3) instead of using the pretrained feature only (K=0) (ablation) | SIB with Conv-4-64 · CIFAR-FS 5-way 1-shot | accuracy: 59.2$\pm$0.5\% -> 68.7$\pm$0.6\% | Helped |
| Add synthetic-gradient transductive updates rather than inductive MAML-like updates (ablation) | SIB on CIFAR-FS with Conv-4-64, trained 1-shot, eta=1e-2, K=3 · CIFAR-FS 5-way 1-shot | accuracy: 59.5$\pm$0.5\% -> 68.6$\pm$0.6\% | Helped |
| Increase synthetic-gradient steps from K=0 to K=3 (ablation) | SIB-Rot on PACS · PACS zero-shot domain adaptation average | accuracy: 87.7\% -> 89.6\% | Helped |
| Vary the number of query images used, from n=3 to n=10 (ablation) | SIB with Conv-4-128 · MiniImageNet 5-way 5-shot | accuracy: 75.91 $\pm$ 0.66\% -> 76.22 $\pm$ 0.66\% | Helped |
| Increase the number of query images beyond n=10 to n=45 (ablation) | SIB with Conv-4-128 · MiniImageNet 5-way 5-shot | accuracy: 76.22 $\pm$ 0.66\% -> 73.27 $\pm$ 0.66\% | Hurt |
| Increase synthetic-gradient steps from K=0 to K=5 (ablation) | SIB with WRN-28-10 · MiniImageNet 5-way 5-shot | accuracy: 77.5$\pm$0.3\% -> 79.2$\pm$0.4\% | Helped |
| Increase synthetic-gradient steps from K=3 to K=5 (ablation) | SIB with WRN-28-10 · CIFAR-FS 5-way 5-shot | accuracy: 85.3$\pm$0.4\% -> 85.3$\pm$0.4\% | No clear effect |
Closest papers by meaning
- Meta-Learning with Shared Amortized Variational Inference (ICML 2020 · similarity 0.89)
- Learning to Balance: Bayesian Meta-Learning for Imbalanced and Out-of-distribution Tasks (ICLR 2020 · similarity 0.88)
- Amortized Bayesian Meta-Learning (ICLR 2019 · similarity 0.87)
- Gradient-EM Bayesian Meta-Learning (NeurIPS 2020 · similarity 0.87)
- Large-Scale Meta-Learning with Continual Trajectory Shifting (ICML 2021 · similarity 0.87)
- Bayesian Uncertainty for Gradient Aggregation in Multi-Task Learning (ICML 2024 · similarity 0.87)
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.