Research map / Probabilistic and scientific modeling / Variational and generative sampling
Learning Discrete Structured Variational Auto-Encoder using Natural Evolution Strategies
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 NES instead of SST for structured latent optimization | NRI · Latent structure recovery: Spanning Tree | Edge F1-Score: 91 ± 3 -> 92 ± 0.2 | Helped |
| Use NES instead of SST for structured latent optimization | NRI · Latent structure recovery: n - 1 Individual Edges | Edge F1-Score: 41 ± 1 -> 44 ± 0.6 | Helped |
| Use NES instead of DPP | VAE dependency parser, projective · UD target domains: GL_CTG, GL_TREEGAL, ID_CSUI, ID_GSD, RU_GSD, RU_TAIGA | UAS: 68.72 -> 68.92; 71.39 -> 71.64; 68.23 -> 68.28; 71.71 -> 71.41; 71.46 -> 71.82; 69.94 -> 70.52 | Mixed |
| Use NES instead of SparseMAP | VAE dependency parser, non-projective · UD target domains: GL_CTG, GL_TREEGAL, ID_CSUI, ID_GSD, RU_GSD, RU_TAIGA | UAS: 68.57 -> 68.67; 70.61 -> 70.98; 68.23 -> 68.28; 70.56 -> 70.47; 70.99 -> 71.01; 69.96 -> 69.96 | Mixed |
| Increase the VAE network size from 25K to 800K parameters | NES-trained FashionMNIST discrete VAE · FashionMNIST | ELBO: -239.40 -> -234.86 | No clear effect |
| Increase the network size by a factor of two while using no more than 50% additional NES samples | NES-trained discrete VAE · FashionMNIST neural network size analysis | NES samples required to reach SST performance: + 0.00%; + 50.00%; + 11.11%; + 0.00%; + 0.00% | No clear effect |
| Bound the training loss with a sufficiently large upper bound (ablation) | NES-trained discrete VAE · FashionMNIST | Negative ELBO | Helped |
| Use a relatively small loss upper bound (ablation) | NES-trained discrete VAE · FashionMNIST | Negative ELBO | Hurt |
| Use a smaller Gaussian perturbation scale σ (ablation) | NES-trained discrete VAE · FashionMNIST | Distance between negative ELBO and its Gaussian approximation | Helped |
| Use NES instead of GSM | Unstructured discrete VAE · KMNIST | Negative ELBO | Helped |
| Use NES instead of GSM | Unstructured discrete VAE · FashionMNIST and Omniglot | Negative ELBO | No clear effect |
| Use NES instead of exact unbiased enumeration | Unstructured discrete VAE · FashionMNIST, KMNIST, and Omniglot | Negative ELBO | No clear effect |
Closest papers by meaning
- Discrete Variational Autoencoding via Policy Search (ICLR 2026 · similarity 0.90)
- Structured Variationally Auto-encoded Optimization (ICML 2018 · similarity 0.88)
- Unbiased learning of deep generative models with structured discrete representations (NeurIPS 2023 · similarity 0.88)
- Direct Optimization through $\arg \max$ for Discrete Variational Auto-Encoder (NeurIPS 2019 · similarity 0.88)
- From Variational to Deterministic Autoencoders (ICLR 2020 · similarity 0.88)
- DVAE++: Discrete Variational Autoencoders with Overlapping Transformations (ICML 2018 · 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.