Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Hypernetwork Approach to Bayesian MAML (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BH-MAML, a Bayesian MAML variant where universal weights are learned point-wise and task-specific posteriors (Gaussian or CNF-based) over weight updates are generated by a hypernetwork.
Problem
Few-shot learning with MAML overfits on small data and lacks uncertainty quantification; existing Bayesian MAML variants use hierarchical distributions over universal weights that complicate optimization.
Approach
Learn universal weights point-wise and place distributions only on task-specialized weights, produced by a hypernetwork conditioned on the support set and universal weights (Gaussian or continuous normalizing flow posteriors), replacing gradient-based inner-loop adaptation and dropping the hierarchical coupling prior.
Key results
- In 1-shot Omniglot→EMNIST, BH-MAML (G) reaches 80.95% vs. 79.07% for HyperMAML and 74.81% for MAML.
- In 5-shot Omniglot→EMNIST, BH-MAML (G) with adaptation reaches 89.76%, comparable to DKT's 90.30%.
- On mini-ImageNet→CUB, BH-MAML performs comparably to MAML and ProtoNet (e.g., 36.90% vs. 34.01% for MAML in 1-shot).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BH-MAML-G instead of the strongest baseline DKT | Cross-domain few-shot classification · Omniglot→EMNIST 1-shot | accuracy: 75.40 -> 80.95 | Helped |
| Use BH-MAML-G instead of the strongest baseline DKT | Cross-domain few-shot classification · Omniglot→EMNIST 5-shot | accuracy: 90.30 -> 89.21 | No clear effect |
| Use BH-MAML-G instead of the strongest baseline DKT | Cross-domain few-shot classification · mini-ImageNet→CUB 1-shot | accuracy: 40.14 -> 36.90 | Hurt |
| Use BH-MAML-G instead of the strongest baseline DKT | Cross-domain few-shot classification · mini-ImageNet→CUB 5-shot | accuracy: 56.40 -> 49.24 | Hurt |
| Replace the Gaussian posterior with a CNF posterior (ablation) | BH-MAML · Omniglot→EMNIST 1-shot | accuracy: 80.95 -> 72.02 | Hurt |
| Replace the Gaussian posterior with a CNF posterior (ablation) | BH-MAML · Omniglot→EMNIST 5-shot | accuracy: 89.21 -> 82.36 | Hurt |
| Replace the Gaussian posterior with a CNF posterior (ablation) | BH-MAML · mini-ImageNet→CUB 1-shot | accuracy: 36.90 -> 33.77 | Hurt |
| Replace the Gaussian posterior with a CNF posterior (ablation) | BH-MAML · mini-ImageNet→CUB 5-shot | accuracy: 49.24 -> 44.09 | Hurt |
| Add the table's ada. variant (ablation) | BH-MAML-G · Omniglot→EMNIST 1-shot | accuracy: 80.95 -> 81.05 | Helped |
| Add the table's ada. variant (ablation) | BH-MAML-G · Omniglot→EMNIST 5-shot | accuracy: 89.21 -> 89.76 | Helped |
| Add the table's ada. variant (ablation) | BH-MAML-G · mini-ImageNet→CUB 1-shot | accuracy: 36.90 -> 37.23 | Helped |
| Add the table's ada. variant (ablation) | BH-MAML-G · mini-ImageNet→CUB 5-shot | accuracy: 49.24 -> 50.79 | Helped |
Surprising findings
- The CNF variant underperforms the Gaussian variant in the reported cross-domain experiments (e.g., 72.02% vs. 80.95% in 1-shot Omniglot→EMNIST).
Limitations
- Results are only on cross-domain few-shot classification; BH-MAML is not clearly superior to baselines in several settings.
- CNF variant underperforms the Gaussian variant on the reported benchmarks.
- Short student abstract; limited experimental detail and no ablations reported.
Experimental setup
Datasets: Omniglot EMNIST mini-ImageNet CUB
Benchmarks: Omniglot to EMNIST mini-ImageNet to CUB
Metrics: classification accuracy
Baselines: Feature Transfer ProtoNet MAML DKT Bayesian MAML HyperMAML
Builds on: MAML Bayesian MAML Amortized Bayesian Meta-Learning Continuous Normalizing Flows HyperMAML
Closest papers by meaning
- How to train your MAML (ICLR 2019 · similarity 0.88)
- Recasting Gradient-Based Meta-Learning as Hierarchical Bayes (ICLR 2018 · similarity 0.87)
- How to Train Your MAML to Excel in Few-Shot Classification (ICLR 2022 · similarity 0.86)
- HyperNet Fields: Efficiently Training Hypernetworks without Ground Truth by Learning Weight Trajectories (CVPR 2025 · similarity 0.84)
- MotherNet: Fast Training and Inference via Hyper-Network Transformers (ICLR 2025 · similarity 0.84)
- Enabling Few-Shot Learning with PID Control: A Layer Adaptive Optimizer (ICML 2024 · similarity 0.83)
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.