Research map / Probabilistic and scientific modeling / Variational and generative sampling
Denoising Implicit Variational Inference
Zitong Wang, Zekun Wu, Longlin Yu, Cheng Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a denoising-based approach to semi-implicit variational inference that addresses the intractable marginal score in ELBO optimization; full text unavailable, so details are limited to the abstract.
Problem
Optimizing the ELBO for semi-implicit variational inference is hard because the marginal score of the semi-implicit distribution is intractable, which existing UIVI-style reparameterized gradient estimators handle imperfectly.
Approach
Frames the intractable marginal score estimation in SIVI as a denoising problem, extending UIVI and related reparameterized ELBO gradient methods; specific mechanism not stated in the abstract provided.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace UIVI reverse-conditional MCMC with a score network trained by denoising score matching | semi-implicit VI · Conditioned diffusion (100-dim) | expected log marginal likelihood: 21.866 -> 75.162 | Improved |
| replace UIVI reverse-conditional MCMC with learned score | semi-implicit VI · Conditioned diffusion (100-dim) | wall-clock time (s): 756 -> 335 | Improved |
| replace UIVI with DIVI on BNN regression | semi-implicit VI · BNN Protein | test RMSE: 5.021 -> 4.493 | Improved |
| replace UIVI with DIVI on BNN regression | semi-implicit VI · BNN Boston | test RMSE: 3.266 -> 2.677 | Improved |
| use DIVI instead of KSIVI on BNN regression | semi-implicit VI · BNN Wine Red and Yacht | test RMSE: Yacht 1.274 (KSIVI) vs 1.358 | Mixed |
| replace UIVI with DIVI on BNN regression (cost) | semi-implicit VI · BNN (six UCI datasets) | avg wall-clock time (s): 842 -> 264 | Improved |
| use DIVI learned score instead of UIVI HMC score estimator (ablation) | fixed DIVI variational checkpoints · X-shaped target, 10,000 steps | squared L2 score error: 899.236 -> 0.331 | Improved |
| use DIVI learned score instead of AISIVI importance-sampling score estimator (ablation) | fixed DIVI variational checkpoints · X-shaped target, 10,000 steps | squared L2 score error: 6.881 -> 0.331 | Improved |
| use DIVI learned score instead of UIVI HMC estimator (ablation) | score evaluation · X-shaped target | score-evaluation time (ms): 204.523 -> 0.134 | Improved |
| use DIVI on two-dimensional targets | semi-implicit VI · 2D toy targets (x_shaped, student_uc, 8_gaussians) | KL-style and Wasserstein metrics | Improved |
Limitations
- Full text unavailable; only the abstract was provided, so experimental details are unknown.
Experimental setup
Baselines: Unbiased Implicit Variational Inference
Builds on: Unbiased Implicit Variational Inference Semi-implicit variational inference
Closest papers by meaning
- Revisiting Unbiased Implicit Variational Inference (ICML 2025 · similarity 0.93)
- Kernel Semi-Implicit Variational Inference (ICML 2024 · similarity 0.92)
- Denoising Diffusion Variational Inference: Diffusion Models as Expressive Variational Posteriors (AAAI 2025 · similarity 0.90)
- Particle Semi-Implicit Variational Inference (NeurIPS 2024 · similarity 0.90)
- EigenVI: score-based variational inference with orthogonal function expansions (NeurIPS 2024 · similarity 0.89)
- Natural Gradient VI: Guarantees for Non-Conjugate Models (NeurIPS 2025 · similarity 0.89)
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.
