Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Lazy Estimation of Variable Importance for Large Neural Networks
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 LazyVI instead of retraining | two-layer ReLU neural network · Binary classification simulation | computation time: 0.6 seconds vs. $7.5$ seconds | Helped |
| Use LazyVI instead of dropout | variable importance estimation · Linear systems with correlated variables | VI estimation accuracy | Helped |
| Use LazyVI instead of retraining | high-dimensional nonlinear regression · 100-dimensional nonlinear regression | normalized error relative to retraining | Helped |
| Use LazyVI instead of dropout | seasonal precipitation VI estimation · CESM-LENS ocean climate indices | VI estimation accuracy relative to retraining | Helped |
| Use LazyVI instead of retraining | three-layer fully connected neural network · High-dimensional seasonal forecasting ROAR | computation time: 1.8s vs. 3.2s | Helped |
| Use LazyVI instead of retraining | three-layer fully connected neural network · High-dimensional seasonal forecasting ROAR | MSE after variable removal | No clear effect |
| Use LazyVI-ES instead of full retraining | convolutional neural network · MNIST ROAR with random ordering | accuracy: nearly identical | No clear effect |
| Use LazyVI-ES instead of full retraining | convolutional neural network · MNIST ROAR with random ordering | computation time: $5\times$ speed-up | Helped |
| Use LazyVI instead of retraining | two-layer neural network Shapley estimation · Sparse high-dimensional logistic-model data | computation time: roughly $5$ times faster | Helped |
| Initialize LazyVI at random rather than at the trained full-model parameters (ablation) | LazyVI · Linear-system simulation | VI estimation accuracy | Hurt |
| Increase the training network width (ablation) | LazyVI · Linear-system simulation | confidence-interval coverage | Helped |
| Increase the training network width (ablation) | LazyVI · Linear-system simulation | computation time | Hurt |
Closest papers by meaning
- Interpretable Predictive Modeling for Climate Variables with Weighted Lasso (AAAI 2019 · similarity 0.87)
- Variable Importance in High-Dimensional Settings Requires Grouping (AAAI 2024 · similarity 0.85)
- Towards a Unified Framework for Uncertainty-aware Nonlinear Variable Selection with Theoretical Guarantees (NeurIPS 2022 · similarity 0.85)
- Total Variation Floodgate for Variable Importance Inference in Classification (ICML 2024 · similarity 0.84)
- Optimizing importance weighting in the presence of sub-population shifts (ICLR 2025 · similarity 0.84)
- High-dimensional Analysis of Knowledge Distillation: Weak-to-Strong Generalization and Scaling Laws (ICLR 2025 · similarity 0.83)
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.