Research map / Probabilistic and scientific modeling / Variational and generative sampling
Reparameterization Gradient for Non-differentiable Models
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 proposed estimator instead of Score | stochastic variational inference · temperature, textmsg, influenza | gradient variance: Score variance ratios are 1; Ours is below 1 in all reported settings | Helped |
| Add boundary contribution estimates to the reparameterization gradient | Ours · textmsg, stepsize=0.001 | average gradient variance: 2.91 × 10^{-2} -> 2.77 × 10^{-2} | Helped |
| Add boundary contribution estimates to the reparameterization gradient | Ours · textmsg, stepsize=0.001 | gradient norm variance: 2.92 × 10^{-2} -> 2.46 × 10^{-2} | Helped |
| Add boundary contribution estimates to the reparameterization gradient | Ours · temperature, stepsize=0.001 | average gradient variance: 4.45 × 10^{-9} -> 1.85 × 10^{-6} | Hurt |
| Add boundary contribution estimates to the reparameterization gradient | Ours · influenza, stepsize=0.001 | average gradient variance: 4.38 × 10^{-3} -> 4.89 × 10^{-3} | Hurt |
| Use Ours rather than Reparam | stochastic variational inference · temperature, textmsg, influenza | ELBO optimization | Mixed |
| Use Reparam rather than Score | stochastic variational inference · temperature, textmsg, influenza | ELBO optimization | Helped |
| Increase the Adam stepsize to 0.01 | Score · temperature and textmsg | optimization stability | Hurt |
| Use the biased Reparam estimator without boundary contributions | Reparam · temperature | ELBO optimization | Hurt |
| Use Ours instead of Reparam | stochastic variational inference · temperature, textmsg, influenza, N=1 | computation time per iteration: 46.1 -> 79.2 ms; 15.4 -> 24.9 ms; 251.4 -> 269.8 ms | Hurt |
| Increase Monte Carlo sample count | all gradient estimators · temperature, textmsg, influenza | convergence speed and stability | Helped |
Closest papers by meaning
- The Generalized Reparameterization Gradient (NeurIPS 2016 · similarity 0.93)
- Approximation Based Variance Reduction for Reparameterization Gradients (NeurIPS 2020 · similarity 0.92)
- Reducing Reparameterization Gradient Variance (NeurIPS 2017 · similarity 0.92)
- Natural Gradient VI: Guarantees for Non-Conjugate Models (NeurIPS 2025 · similarity 0.89)
- Rao-Blackwellised Reparameterisation Gradients (NeurIPS 2025 · similarity 0.89)
- Understanding Stochastic Natural Gradient Variational Inference (ICML 2024 · similarity 0.88)
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.