Research map / Optimization / Implicit bias and generalization
What You See is What You Get: Principled Deep Learning via Distributional Generalization
Bogdan Kulynych, Yao-Yuan Yang, Yaodong Yu, Jarosław Błasiok, Preetum Nakkiran
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 |
|---|---|---|---|
| replace l2 regularization with Gaussian gradient noise (ablation) | IS-SGD · CivilComments | worst-group accuracy: 64.3 -> 71.9 | Improved |
| replace l2 regularization with Gaussian gradient noise (ablation) | IS-SGD · iNaturalist2017 | worst-group accuracy: 70.6 -> 71.0 | Improved |
| replace l2 regularization with Gaussian gradient noise (ablation) | IS-SGD · CelebA | worst-group accuracy: 82.4 -> 84.9 | Improved |
| replace l2 regularization with Gaussian gradient noise (ablation) | IW-SGD · CelebA | worst-group accuracy: 89.0 -> 88.5 | No clear effect |
| replace l2 regularization with Gaussian gradient noise (ablation) | IW-SGD · MultiNLI | worst-group accuracy: 68.1 -> 69.7 | Mixed |
| replace l2 regularization with Gaussian gradient noise (ablation) | gDRO · iNaturalist2017 | worst-group accuracy: 67.3 -> 56.4 | Hurt |
| replace l2 regularization with Gaussian gradient noise (ablation) | gDRO · CelebA | worst-group accuracy: 84.5 -> 83.3 | Hurt |
| replace l2 regularization with Gaussian gradient noise (ablation) | gDRO · CivilComments | worst-group accuracy: 67.3 -> 71.3 | Improved |
| add DP (clipping and noise) to importance sampling (ablation) | IS-SGD · CelebA | worst-group accuracy: 82.4 -> 86.0 | Improved |
| add DP (clipping and noise) to importance sampling (ablation) | IS-SGD · MultiNLI | worst-group accuracy: 70.4 -> 72.3 | Improved |
| add DP (clipping and noise) to importance sampling (ablation) | IS-SGD · iNaturalist2017 | worst-group accuracy: 70.6 -> 51.4 | Hurt |
| use importance sampling instead of DP-SGD uniform sampling (ablation) | DP-SGD · CelebA, epsilon 2 to 12 | test accuracy: around 8 p.p. lower | Hurt |
Closest papers by meaning
- Position: Understanding LLMs Requires More Than Statistical Generalization (ICML 2024 · similarity 0.85)
- Interpretability Illusions in the Generalization of Simplified Models (ICML 2024 · similarity 0.84)
- Relating Misfit to Gain in Weak-to-Strong Generalization Beyond the Squared Loss (ICML 2025 · similarity 0.84)
- To Grok or not to Grok: Disentangling Generalization and Memorization on Corrupted Algorithmic Datasets (ICLR 2024 · similarity 0.84)
- Differentially private optimization for non-decomposable objective functions (ICLR 2025 · similarity 0.84)
- Position: Deep Learning is Not So Mysterious or Different (ICML 2025 · similarity 0.84)
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.
