Research map / Probabilistic and scientific modeling / Variational and generative sampling
Exemplar VAE: Linking Generative Models, Nearest Neighbor Retrieval, and Data Augmentation
Sajad Norouzi, David J Fleet, Mohammad Norouzi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 an exemplar-based latent mixture prior instead of a Gaussian prior | VAE · Dynamic MNIST | test log-likelihood lower bound (nats): -84.45 ± 0.12 -> -82.09 ± 0.18 | Improved |
| Use an exemplar-based latent mixture prior instead of a Gaussian prior | VAE · Fashion MNIST | test log-likelihood lower bound (nats): -228.70 ± 0.15 -> -226.75 ± 0.07 | Improved |
| Use an exemplar-based latent mixture prior instead of a Gaussian prior | VAE · Omniglot | test log-likelihood lower bound (nats): -108.34 ± 0.06 -> -105.22 ± 0.18 | Improved |
| Use an exemplar-based latent mixture prior instead of a Gaussian prior | PixelSNAIL · Dynamic MNIST | test log-likelihood lower bound (nats): -78.20 ± 0.02 -> -77.95 { ± 0.01} | Hurt |
| Add Exemplar VAE-generated examples to label-smoothed classifier training | MLP classifier · Permutation-invariant MNIST, 2×8192 hidden units | test error (%): 1.17 -> 0.69 | Improved |
| Add Exemplar VAE-generated examples to label-smoothed classifier training | MLP classifier · Permutation-invariant Fashion MNIST, 2×8192 hidden units | test error (%): 8.56 -> 8.16 | Improved |
| Add leave-one-out exemplar training (ablation) | Exemplar VAE · MNIST | test log-likelihood lower bound (nats): -101.33 -> -82.35 | Improved |
| Add leave-one-out exemplar training (ablation) | Exemplar VAE · Omniglot | test log-likelihood lower bound (nats): -139.12 -> -105.80 | Improved |
| Subsample exemplars to M/N=0.5 rather than using M/N=1 (ablation) | Exemplar VAE with LOO · MNIST | test log-likelihood lower bound (nats): -82.35 -> -82.09 | Improved |
| Subsample exemplars to M/N=0.2 rather than using M/N=1 (ablation) | Exemplar VAE with LOO · Omniglot | test log-likelihood lower bound (nats): -105.80 -> -104.95 | Improved |
Later papers that compared with this method
1 later accepted paper reports Exemplar VAE (25000) as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- ByPE-VAE: Bayesian Pseudocoresets Exemplar VAE (NeurIPS 2021 · better on 3 of 3 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Retrieval-Augmented Diffusion Models (NeurIPS 2022 · similarity 0.87)
- VAE Approximation Error: ELBO and Exponential Families (ICLR 2022 · similarity 0.86)
- Perceptual Generative Autoencoders (ICML 2020 · similarity 0.86)
- Trading Information between Latents in Hierarchical Variational Autoencoders (ICLR 2023 · similarity 0.86)
- EQ-VAE: Equivariance Regularized Latent Space for Improved Generative Image Modeling (ICML 2025 · similarity 0.86)
- Generalization in VAE and Diffusion Models: A Unified Information-Theoretic Analysis (ICLR 2025 · similarity 0.85)
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.
