Research map / Probabilistic and scientific modeling / Variational and generative sampling
Bias Correction of Learned Generative Models using Likelihood-Free Importance Weighting
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 |
|---|---|---|---|
| Apply self-normalized LFIW to PixelCNN++ samples | PixelCNN++ · CIFAR-10 | IS: 5.16 ± 0.0117 -> 6.68 ± 0.0773 | Helped |
| Apply self-normalized LFIW to PixelCNN++ samples | PixelCNN++ · CIFAR-10 | FID: 58.70 ± 0.0506 -> 55.83 ± 0.9695 | Helped |
| Apply self-normalized LFIW to SNGAN samples | SNGAN · CIFAR-10 | KID: 0.094 ± 0.0002 -> 0.073 ± 0.0004 | Helped |
| Importance-weight generated points in the real-plus-generated training set | DAGAN augmentation classifier · Omniglot | accuracy: 0.6600 ± 0.0040 -> 0.6818 ± 0.0022 | Helped |
| Use generated training points without LFIW (ablation) | DAGAN augmentation classifier · Omniglot | accuracy: 0.6603 ± 0.0012 -> 0.6600 ± 0.0040 | No clear effect |
| Weight simulated trajectories with LFIW | Learned dynamics model · Swimmer | policy return: 100.4 ± 3.2 -> 25.7 ± 3.1 | Mixed |
| Weight simulated trajectories with LFIW | Learned dynamics model · HalfCheetah | policy return: 204.0 ± 0.8 -> 217.8 ± 4.0 | Helped |
| Weight simulated trajectories with LFIW | Learned dynamics model · HumanoidStandup | policy return: 8417 ± 28 -> 9372 ± 375 | Helped |
| Use flattening with α=0.25 instead of self-normalized weighting (ablation) | SNGAN LFIW · CIFAR-10 Inception activations | MSE: 0.0016 ± 0.00023 -> 0.0011 ± 8e-05 | Helped |
| Use flattening with α=0.25 instead of self-normalized weighting (ablation) | PixelCNN++ LFIW · CIFAR-10 Inception activations | MSE: 0.0046 ± 0.00031 -> 0.0175 ± 0.00138 | Hurt |
| Clip importance weights at lower bound β=1.0 (ablation) | SNGAN LFIW · CIFAR-10 Inception activations | MSE: 0.0016 ± 0.00023 -> 0.0170 ± 0.00171 | Hurt |
| Clip importance weights at lower bound β=0.1 (ablation) | PixelCNN++ LFIW · CIFAR-10 Inception activations | MSE: 0.0046 ± 0.00031 -> 0.0215 ± 0.00149 | Hurt |
Closest papers by meaning
- Fair Generative Modeling via Weak Supervision (ICML 2020 · similarity 0.88)
- Bias and Generalization in Deep Generative Models: An Empirical Study (NeurIPS 2018 · similarity 0.88)
- Understanding Deep Generative Models With Generalized Empirical Likelihoods (CVPR 2023 · similarity 0.87)
- On the Quantitative Analysis of Decoder-Based Generative Models (ICLR 2017 · similarity 0.87)
- Optimizing importance weighting in the presence of sub-population shifts (ICLR 2025 · similarity 0.87)
- Statistically Optimal Generative Modeling with Maximum Deviation from the Empirical Distribution (ICML 2024 · similarity 0.86)
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.