Research map / Optimization / Implicit bias and generalization
Time-Independent Information-Theoretic Generalization Bounds for SGLD
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 |
|---|---|---|---|
| Use time-evolving KL divergence between outputs trained on different datasets instead of bounding mutual information over the full iterate sequence | SGLD · Expected generalization error with surrogate loss | expected generalization error bound: \mathcal{O}(\sqrt{\eta t / n}) -> \mathcal{O}(\sqrt{(\eta t\wedge 1)/n)} | Helped |
| Use the proposed time-independent information-theoretic bound rather than the reflection-coupling stability bound | SGLD · Expected generalization error with surrogate loss | dependence on step size: \mathcal{O}((\eta t\wedge 1)(\sqrt{\eta^{-1}}/n+\sqrt{\eta})) -> \mathcal{O}(\sqrt{(\eta t\wedge 1)/n)} | Helped |
| Evaluate generalization with the same loss used for training, using the derived sub-exponential tail property | SGLD · Expected generalization error with same training and evaluation loss | availability of an information-theoretic bound | Helped |
| Combine the same-loss generalization bound with existing non-convex optimization error bounds | SGLD · Excess risk | excess risk bound | Helped |
| Use Gaussian noise in the noisy iterative algorithm | Proposed analysis · SGLD generalization analysis | scope of applicable noise distributions | Hurt |
Closest papers by meaning
- Time-independent Generalization Bounds for SGLD in Non-convex Settings (NeurIPS 2021 · similarity 0.96)
- Information-Theoretic Generalization Bounds for SGLD via Data-Dependent Estimates (NeurIPS 2019 · similarity 0.95)
- Optimizing Information-theoretical Generalization Bound via Anisotropic Noise of SGLD (NeurIPS 2021 · similarity 0.95)
- Generalization of noisy SGD in unbounded non-convex settings (ICML 2025 · similarity 0.94)
- Temperature is All You Need for Generalization in Langevin Dynamics and other Markov Processes (NeurIPS 2025 · similarity 0.90)
- Exactly Tight Information-theoretic Generalization Bounds via Binary Jensen-Shannon Divergence (ICML 2025 · similarity 0.89)
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.