Research map / Probabilistic and scientific modeling / Variational and generative sampling
Constraining Variational Inference with Geometric Jensen-Shannon Divergence
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 dual skew-geometric JS divergence with a tuned α | VAE · MNIST, Fashion-MNIST, dSprites, and Chairs | reconstruction error (MSE): KL(q(z|x)||p(z)): 8.46, 11.98, 13.55, 12.27; JS^G_α*: 7.34, 9.58, 4.97, 11.64 | Helped |
| Use JS^G_α with a tuned α instead of KL(q(z|x)||p(z)) | VAE · MNIST, Fashion-MNIST, dSprites, and Chairs | reconstruction error (MSE): 8.46 -> 7.52; 11.98 -> 10.04; 13.55 -> 5.54; 12.27 -> 11.95 | Helped |
| Use β=0.25 instead of the standard KL baseline (ablation) | β-VAE · MNIST | reconstruction error (MSE): 8.46 -> 8.09 | Helped |
| Use β=0.25 instead of the standard KL baseline (ablation) | β-VAE · Fashion-MNIST | reconstruction error (MSE): 11.98 -> 9.07 | Helped |
| Use β=0.25 instead of the standard KL baseline (ablation) | β-VAE · dSprites | reconstruction error (MSE): 13.55 -> 10.39 | Helped |
| Use β=0.25 instead of the standard KL baseline (ablation) | β-VAE · Chairs | reconstruction error (MSE): 12.27 -> 14.09 | Hurt |
| Set α=0.5 rather than the lowest-error skew reported for JS^G_α (ablation) | JS^G_α VAE · MNIST, Fashion-MNIST, dSprites, and Chairs | reconstruction error (MSE): 7.52 -> 9.87; 10.04 -> 11.29; 5.54 -> 9.89; 11.95 -> 13.57 | Hurt |
| Increase α above 0.5 in JS^G_α (ablation) | JS^G_α VAE · MNIST and Fashion-MNIST | reconstruction error (MSE): MNIST: 9.87 -> 11.87; Fashion-MNIST: 11.29 -> 13.21 | Hurt |
| Use JS^G_α* for higher generative performance rather than its low-error settings | VAE · MNIST and Fashion-MNIST | model evidence | Mixed |
Closest papers by meaning
- A Geometric Perspective on Variational Autoencoders (NeurIPS 2022 · similarity 0.90)
- MAE: Mutual Posterior-Divergence Regularization for Variational AutoEncoders (ICLR 2019 · similarity 0.89)
- Variational Autoencoders with Riemannian Brownian Motion Priors (ICML 2020 · similarity 0.89)
- $t^3$-Variational Autoencoder: Learning Heavy-tailed Data with Student's t and Power Divergence (ICLR 2024 · similarity 0.89)
- On Implicit Regularization in $β$-VAEs (ICML 2020 · similarity 0.89)
- Asymptotics of Alpha-Divergence Variational Inference Algorithms with Exponential Families (NeurIPS 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.