Research map / Learning theory and algorithms / Kernels and regression theory
Meta-Learning by Adjusting Priors Based on Extended PAC-Bayes Theory
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 |
|---|---|---|---|
| learn an experience-dependent prior from all observed tasks using MLAP-M | CNN on permuted-label tasks · Permuted Labels | test error (%): 2.27 -> 0.908 | Helped |
| learn an experience-dependent prior from all observed tasks using MLAP-M | fully connected network · Permuted Pixels (200 swaps) | test error (%): 7.92 -> 3.4 | Helped |
| replace the standard scratch learning initialization with a single-task warm-start | learning on permuted-label tasks · Permuted Labels | test error (%): 2.82 -> 1.07 | Helped |
| replace the standard scratch learning initialization with a single-task warm-start | learning on permuted-pixel tasks · Permuted Pixels (200 swaps) | test error (%): 7.65 -> 7.95 | No clear effect |
| use a hand-tailored frozen-layer transfer method | learning on permuted-label tasks · Permuted Labels | test error (%): 2.27 -> 0.69 | Helped |
| use a hand-tailored frozen-layer transfer method | learning on permuted-pixel tasks · Permuted Pixels (200 swaps) | test error (%): 3.4 -> 6.57 | Hurt |
| use MLAP-S with one or two meta-training tasks instead of scratch learning | CNN on permuted-label tasks · Permuted Labels | test error (%) | Hurt |
| increase the number of meta-training tasks | MLAP-S · Permuted Labels and Permuted Pixels | test error | Helped |
| use Seeger's single-task bound as the objective instead of the McAllester bound (ablation) | MLAP · Permuted Labels | test error (%): 0.908 -> 0.75 | Helped |
| use Seeger's single-task bound as the objective instead of the McAllester bound (ablation) | MLAP · Permuted Pixels (200 swaps) | test error (%): 3.4 -> 3.54 | Hurt |
| replace the PAC-Bayes objective with a hierarchical variational Bayes objective (ablation) | MLAP · Permuted Labels | test error (%): 0.908 -> 0.85 | Helped |
| replace the PAC-Bayes objective with a hierarchical variational Bayes objective (ablation) | MLAP · Permuted Pixels (200 swaps) | test error (%): 3.4 -> 3.52 | Hurt |
Closest papers by meaning
- More Flexible PAC-Bayesian Meta-Learning by Learning Learning Algorithms (ICML 2024 · similarity 0.94)
- Fast-Rate PAC-Bayesian Generalization Bounds for Meta-Learning (ICML 2022 · similarity 0.92)
- A Unified View on PAC-Bayes Bounds for Meta-Learning (ICML 2022 · similarity 0.92)
- Fast Rate Bounds for Multi-Task and Meta-Learning with Different Sample Sizes (NeurIPS 2025 · similarity 0.89)
- Generalization Bounds via Meta-Learned Model Representations: PAC-Bayes and Sample Compression Hypernetworks (ICML 2025 · similarity 0.88)
- Recursive PAC-Bayes: A Frequentist Approach to Sequential Prior Updates with No Information Loss (NeurIPS 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.