Research map / Probabilistic and scientific modeling / Variational and generative sampling
An Instability in Variational Inference for Topic 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 naive mean field instead of the information-theoretically uninformative posterior | LDA · LDA synthetic, β between β_inst and β_Bayes | topic/weight correlation with ground truth | Mixed |
| Use naive mean field instead of TAP | LDA · LDA synthetic, β between β_inst and β_Bayes | distance from the uninformative fixed point | Hurt |
| Replace the nominally 0.9 credible interval with naive mean-field intervals at β=2 | LDA synthetic, n=d=5000 · LDA synthetic, β=2 | coverage: 0.9 -> 0.87 | Hurt |
| Replace the nominally 0.9 credible interval with naive mean-field intervals at β=4.1 | LDA synthetic, n=d=5000 · LDA synthetic, β=4.1 | coverage: 0.9 -> 0.65 | Hurt |
| Replace the nominally 0.9 credible interval with naive mean-field intervals at β=6 | LDA synthetic, n=d=5000 · LDA synthetic, β=6 | coverage: 0.9 -> 0.51 | Hurt |
| Use TAP free energy and its AMP iteration instead of naive mean-field free energy | LDA · LDA synthetic, β between β_inst and β_spect | distance from the uninformative fixed point | Helped |
| Include the TAP correction to the naive mean-field free energy | Z₂-synchronization · Z₂-synchronization, λ<1 | local-minimum stability of the uninformative point: 1-2λ -> (1-λ)^2 | Helped |
Closest papers by meaning
- A Discrete Variational Recurrent Topic Model without the Reparametrization Trick (NeurIPS 2020 · similarity 0.90)
- Autoencoding Variational Inference For Topic Models (ICLR 2017 · similarity 0.89)
- Provable Algorithms for Inference in Topic Models (ICML 2016 · similarity 0.89)
- Discovering Discrete Latent Topics with Neural Variational Inference (ICML 2017 · similarity 0.87)
- TreeVI: Reparameterizable Tree-structured Variational Inference for Instance-level Correlation Capturing (NeurIPS 2024 · similarity 0.85)
- Latent Thought Models with Variational Bayes Inference-Time Computation (ICML 2025 · 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.