Research map / Learning theory and algorithms / Prediction and decision losses
On Optimal Generalizability in Parametric Learning
Ahmad Beirami, Meisam Razaviyayn, Shahin Shahrampour, Vahid Tarokh
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 |
|---|---|---|---|
| Approximate each leave-one-out fit using a Hessian-based parameter update instead of retraining for every held-out sample | logistic regression · MNIST (digits 2 and 3) | out-of-sample loss estimate: 0.1056 (0.0179) -> 0.1059 (0.0179) | No clear effect |
| Use the influence-function approximation instead of the proposed ALOOCV | logistic regression · MNIST (digits 2 and 3) | out-of-sample loss estimate: 0.1056 (0.0179) vs 0.0734 (0.0100) | Mixed |
| Reduce regularization from λ=0.8333 to λ=0.0521 | logistic regression · MNIST (digits 2 and 3) | LOOCV: 0.1085 (0.0214) -> 0.1549 (0.0411) | Hurt |
| Use ALOOCV in place of repeatedly solving the leave-one-out training problems | cross-validation estimation · CIFAR-10 (bird vs cat) | runtime per sample: ≈60s -> ≈1s | Improved |
| Tune regularization parameters using the approximate cross-validation gradient | elastic net regression · synthetic regression | loss | Improved |
| Set feature-specific regularization parameters through the proposed gradient descent algorithm | linear regression with 50 feature-specific regularizers · synthetic regression (n=150, p=50) | out-of-sample loss | Improved |
Closest papers by meaning
- Approximate Leave-One-Out for Fast Parameter Tuning in High Dimensions (ICML 2018 · similarity 0.88)
- Safe Grid Search with Optimal Complexity (ICML 2019 · similarity 0.85)
- Certified Minimax Unlearning with Generalization Rates and Deletion Capacity (NeurIPS 2023 · similarity 0.85)
- Hessian-Free Online Certified Unlearning (ICLR 2025 · similarity 0.84)
- Machine Unlearning under Overparameterization (NeurIPS 2025 · similarity 0.84)
- Loss Landscape Characterization of Neural Networks without Over-Parametrization (NeurIPS 2024 · 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.
