Research map / Probabilistic and scientific modeling / Variational and generative sampling
Bias and Generalization in Deep Generative Models: An Empirical Study
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 |
|---|---|---|---|
| Train on a single numerosity value | WGAN-GP and VAE · Colored dots and CLEVR | generated numerosity distribution | Mixed |
| Train on distant feature modes; the response behaves like convolution with the single-mode response | WGAN-GP and VAE · Pie dataset | agreement with convolution prediction | Helped |
| Train on close feature modes | WGAN-GP and VAE · Pie dataset | probability assigned to the mean feature value | Helped |
| Vary distributions of other features while holding the feature under study fixed | WGAN-GP and VAE · Pie dataset | feature-wise distribution stability | No clear effect |
| Increase randomness in other feature dimensions | WGAN-GP and VAE · Pie dataset | variance of the learned feature distribution | Mixed |
| Increase the number of training combinations | GANs and VAEs · Pie dataset and Three MNIST | precision | Hurt |
| Increase the number of training combinations | GANs and VAEs · Pie dataset and Three MNIST | recall | No clear effect |
| Train with more existing shape-color combinations | WGAN-GP · CLEVR | novel combination generation | Helped |
| Increase network capacity (ablation) | WGAN · Pie dataset, feature combinations | precision-recall curve: 3M parameters to 24M | No clear effect |
| Change the number of training examples (ablation) | GANs and VAEs · Pie dataset and Three MNIST | precision-recall behavior: half or twice as many training examples | No clear effect |
| Replace convolutional architecture and Gaussian latent noise with fully connected architecture and Bernoulli latent noise (ablation) | WGAN-GP · Probed feature-generalization tasks | qualitative generalization patterns | No clear effect |
| Replace the convolutional VAE with a fully connected β-VAE using coefficient 3 and a larger learning rate (ablation) | VAE · Probed feature-generalization tasks | qualitative generalization patterns | No clear effect |
Closest papers by meaning
- Can Biases in ImageNet Models Explain Generalization? (CVPR 2024 · similarity 0.89)
- Would Deep Generative Models Amplify Bias in Future Models? (CVPR 2024 · similarity 0.89)
- Statistically Optimal Generative Modeling with Maximum Deviation from the Empirical Distribution (ICML 2024 · similarity 0.87)
- A Likelihood Based Approach to Distribution Regression Using Conditional Deep Generative Models (ICML 2025 · similarity 0.87)
- Data Distributional Properties As Inductive Bias for Systematic Generalization (CVPR 2025 · similarity 0.87)
- Low-Biased General Annotated Dataset Generation (CVPR 2025 · similarity 0.87)
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.