Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
The Memory-Perturbation Equation: Understanding Model's Sensitivity to Data
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 the iBLR algorithm's diagonal preconditioner for the LOO estimate instead of an ad hoc Hessian approximation | Sensitivity-based LOO estimation during training · FMNIST | faithfulness to test NLL | Helped |
| Use a diagonal-GGN Hessian approximation for the LOO estimate | SGD sensitivity-based LOO estimation during training · FMNIST | faithfulness to test NLL | Hurt |
| Replace the diagonal-GGN approximation with a K-FAC Hessian approximation | SGD sensitivity-based LOO estimation during training · FMNIST | faithfulness to test NLL | Helped |
| Use ad hoc Hessian-based sensitivity estimates during training | SGD and Adam · Training generalization prediction | faithfulness to test NLL | Mixed |
| Estimate output deviation using prediction error and prediction variance | MPE sensitivity estimation · Three datasets, architecture varies by dataset | correlation with true prediction deviations | Helped |
| Use leave-group-out estimates to predict the effect of removing a whole class | MPE · FMNIST | correlation with test NLL | Helped |
| Estimate LOO-CV error from sensitivity during training | MPE · ResNet-20 on CIFAR10 | faithfulness to test NLL | Helped |
Closest papers by meaning
- TPV: Parameter Perturbations Through the Lens of Test Prediction Variance (ICML 2026 · similarity 0.86)
- Forget Sharpness: Perturbed Forgetting of Model Biases Within SAM Dynamics (ICML 2024 · similarity 0.86)
- Model Sensitivity Aware Continual Learning (NeurIPS 2024 · similarity 0.85)
- Stability Evaluation through Distributional Perturbation Analysis (ICML 2024 · similarity 0.85)
- Improving Perturbation-based Explanations by Understanding the Role of Uncertainty Calibration (NeurIPS 2025 · similarity 0.85)
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 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.