Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Gaussian Process Modeling of Approximate Inference Errors for Variational Autoencoders
Check 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 |
|---|---|---|---|
| Model posterior-inference discrepancy with GP noise processes and marginalize the GP posterior | VAE encoder · MNIST | test log-likelihood: 696.51.5 vs 699.42.1 (dim(z)=10); 944.32.8 vs 945.01.7 (dim(z)=20); 1212.93.2 vs 1204.34.0 (dim(z)=50) | Mixed |
| Model posterior-inference discrepancy with GP noise processes and marginalize the GP posterior | VAE encoder · OMNIGLOT | test log-likelihood: 354.81.5 vs 351.11.7 (dim(z)=10); 516.32.1 vs 509.11.4 (dim(z)=20); 821.84.4 vs 732.03.1 (dim(z)=50) | Helped |
| Model posterior-inference discrepancy with GP noise processes and marginalize the GP posterior | VAE encoder · SVHN | test log-likelihood: 3417.611.8 vs 3392.012.6 (dim(z)=10); 4133.213.3 vs 4086.910.9 (dim(z)=20); 5410.019.5 vs 5405.110.4 (dim(z)=50) | Helped |
| Model posterior-inference discrepancy with GP noise processes and marginalize the GP posterior | VAE encoder · CelebA | test log-likelihood: 9843.423.8 vs 9849.512.1 (dim(z)=10); 12184.022.7 vs 12194.211.5 (dim(z)=20); 15370.927.6 vs 15366.212.7 (dim(z)=50) | Mixed |
| Use GP posterior-marginalized encoder instead of test-time SVI gradient updates | VAE · MNIST, dim(z)=50, batch size 128 | inference time (ms/batch): 60.8 -> 9.9 | Helped |
| Increase the number of test-time SVI gradient updates | Semi-amortized VAE · MNIST, dim(z)=50, batch size 128 | inference time (ms/batch): 9.7 -> 60.8 | Hurt |
| Vary the number of SVI gradient steps | Semi-amortized VAE · multiple benchmark datasets | test log-likelihood | Mixed |
| Increase flow count in the IAF or Householder-flow encoder | VAE · multiple benchmark datasets | test log-likelihood | Mixed |
| Increase the number of mixture components in the mixture encoder | VAE · multiple benchmark datasets | test log-likelihood | Hurt |
| Use recursive mixture estimation rather than a blind mixture encoder | VAE · multiple benchmark datasets | test log-likelihood | Mixed |
| Use GP posterior variance as an indicator of posterior approximation difficulty | GPVAE · MNIST, 2D latent space | alignment with true-posterior non-Gaussianity | Helped |
Closest papers by meaning
- Amortized Inference Regularization (NeurIPS 2018 · similarity 0.93)
- Semi-Amortized Variational Autoencoders (ICML 2018 · similarity 0.92)
- Lagging Inference Networks and Posterior Collapse in Variational Autoencoders (ICLR 2019 · similarity 0.92)
- Bayesian Basis Function Approximation for Scalable Gaussian Process Priors in Deep Generative Models (ICML 2025 · similarity 0.91)
- Neighbour-Driven Gaussian Process Variational Autoencoders for Scalable Structured Latent Modelling (ICML 2025 · similarity 0.90)
- Amortized Structured Stochastic Variational Inference for Gaussian Process Latent Variable Models (NeurIPS 2026 · 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.