Research map / Learning theory and algorithms / Kernels and regression theory
Evaluated CMI Bounds for Meta Learning: Tightness and Expressiveness
Fredrik Hellström, Giuseppe Durisi
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 |
|---|---|---|---|
| Use the two-step e-CMI derivation instead of earlier mutual-information meta-learning bounds | Two-step square-root bound · Average meta-learning generalization bound | bound tightness | Improved |
| Replace the two-step derivation with a one-step derivation over joint task and sample selection | One-step square-root bound · Average meta-learning generalization bound | decay rate in n and m | Improved |
| Use disintegrated CMI terms instead of mutual information with the full training data | e-CMI bounds · Deterministic learners with continuous outputs | bound finiteness | Improved |
| Use binary KL divergence bounds instead of square-root bounds | e-CMI bounds · Low training loss regime | sample-size dependence | Improved |
| Assume zero training loss (interpolation) in the binary KL bound | Binary KL bound · Interpolating regime | rate of decay | Improved |
| Use unobserved-task training loss as auxiliary loss instead of observed-task test loss (ablation) | Two-step square-root bound · Representation learning with Natarajan and VC classes | decoupling of complexity terms | Improved |
| Apply the one-step bound to Natarajan and VC dimension classes in the interpolating setting | e-CMI bound · Representation learning, zero training loss | generalization rate | Improved |
| Convert average bounds to high-probability bounds via probabilistic KL bounds | Two-step and one-step bounds · Representation learning | recovery of classical rates | Improved |
| Specialize to empirical risk minimizers with a task-diversity assumption | One-step bound · Excess risk on a fixed target task | excess risk scaling | Improved |
| Use the two-step bound versus the one-step bound for the combined complexity scaling | High-probability e-CMI bounds · Representation learning | scaling in n and m | Mixed |
Closest papers by meaning
- Tighter CMI-Based Generalization Bounds via Stochastic Projection and Quantization (NeurIPS 2025 · similarity 0.90)
- Certification-Enhanced Generalization Bounds (NeurIPS 2026 · similarity 0.86)
- Slicing Mutual Information Generalization Bounds for Neural Networks (ICML 2024 · similarity 0.86)
- Generalization Bounds via Meta-Learned Model Representations: PAC-Bayes and Sample Compression Hypernetworks (ICML 2025 · similarity 0.86)
- Exactly Tight Information-theoretic Generalization Bounds via Binary Jensen-Shannon Divergence (ICML 2025 · similarity 0.85)
- Fast Rate Bounds for Multi-Task and Meta-Learning with Different Sample Sizes (NeurIPS 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.
