Research map / Optimization / Optimizers
Reducing Reparameterization Gradient Variance
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 HVP+Local control variate instead of the pure Monte Carlo estimator | MCVI with the frisk model · frisk, early, mid, and late optimization; L=3 and L=10 | gradient norm variance (% of MC): 100.000 -> 0.900; 100.000 -> 0.078; 100.000 -> 0.014; 1.037; 0.071; 0.022 | Helped |
| Replace the Hessian-vector product with a diagonal Hessian approximation (ablation) | HVP+Local for MCVI on frisk · frisk, early, mid, and late optimization; L=3 and L=10 | gradient norm variance (% of MC): 22.090; 22.804; 38.725; 21.684; 21.260; 53.777 | Hurt |
| Use the exact full Hessian rather than the local approximation for the scale-gradient control variate (ablation) | HVP+Local for MCVI on frisk · frisk, later optimization | log-scale gradient variance: 98.523; 99.811 -> 1.686; 0.431 | Helped |
| Use HVP+Local rather than pure MC gradients | MCVI optimization on frisk · frisk | ELBO convergence and stability | Helped |
| Use HVP+Local rather than pure MC gradients | MCVI optimization on the Bayesian neural network · bnn, wine | ELBO versus wall-clock time | Helped |
Closest papers by meaning
- Approximation Based Variance Reduction for Reparameterization Gradients (NeurIPS 2020 · similarity 0.92)
- Reparameterization Gradient for Non-differentiable Models (NeurIPS 2018 · similarity 0.92)
- The Generalized Reparameterization Gradient (NeurIPS 2016 · similarity 0.92)
- Evaluating the Variance of Likelihood-Ratio Gradient Estimators (ICML 2017 · similarity 0.89)
- Rao-Blackwellised Reparameterisation Gradients (NeurIPS 2025 · similarity 0.87)
- Reparameterization invariance in approximate Bayesian inference (NeurIPS 2024 · similarity 0.86)
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.