Research map / Probabilistic and scientific modeling / Variational and generative sampling
DisARM: An Antithetic Gradient Estimator for Binary Latent Variables
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 |
|---|---|---|---|
| Integrate out the Logistic augmentation randomness with antithetic conditioning | ARM · Bernoulli latent-variable gradient estimation | Gradient-estimator variance | Helped |
| Use DisARM instead of ARM | Bernoulli VAE, linear models · Dynamic MNIST, Fashion MNIST, Omniglot | Train ELBO: -117.66 ± 0.04 -> -116.30 ± 0.08; -256.80 ± 0.16 -> -255.97 ± 0.07; -122.45 ± 0.10 -> -121.15 ± 0.12 | Helped |
| Use DisARM instead of REINFORCE LOO | Bernoulli VAE, linear models · Dynamic MNIST, Fashion MNIST, Omniglot | Train ELBO: -116.57 ± 0.15 -> -116.30 ± 0.08; -256.33 ± 0.14 -> -255.97 ± 0.07; -121.66 ± 0.10 -> -121.15 ± 0.12 | Helped |
| Use DisARM instead of RELAX | Bernoulli VAE, linear models · Dynamic MNIST, Fashion MNIST, Omniglot | Train ELBO: -115.93 ± 0.15 -> -116.30 ± 0.08; -255.83 ± 0.03 -> -255.97 ± 0.07; -120.79 ± 0.09 -> -121.15 ± 0.12 | Mixed |
| Use DisARM instead of ARM | Bernoulli VAE, nonlinear models · Dynamic MNIST, Fashion MNIST, Omniglot | Train ELBO: -107.32 ± 0.28 -> -102.56 ± 0.19; -241.30 ± 0.10 -> -237.77 ± 0.08; -118.76 ± 0.05 -> -115.08 ± 0.11 | Helped |
| Use DisARM instead of REINFORCE LOO | Bernoulli VAE, nonlinear models · Dynamic MNIST, Fashion MNIST, Omniglot | Train ELBO: -102.45 ± 0.12 -> -102.56 ± 0.19; -237.66 ± 0.11 -> -237.77 ± 0.08; -115.26 ± 0.15 -> -115.08 ± 0.11 | Mixed |
| Use DisARM instead of RELAX | Bernoulli VAE, nonlinear models · Bernoulli VAE across three datasets | Train ELBO: -102.53 ± 0.15 -> -102.56 ± 0.19; -238.23 ± 0.17 -> -237.77 ± 0.08; -116.56 ± 0.15 -> -115.08 ± 0.11 | Mixed |
| Use DisARM instead of VIMCO at matched computation | Linear Bernoulli VAE · Dynamic MNIST, train multi-sample bound | Multi-sample bound: -115.80 ± 0.08 -> -114.06 ± 0.13; -109.40 ± 0.07 -> -108.61 ± 0.08 | Helped |
| Use DisARM instead of VIMCO | Nonlinear Bernoulli VAE · Fashion MNIST, train multi-sample bound | Multi-sample bound: -236.41 ± 0.10 -> -236.91 ± 0.10 | Hurt |
| Use DisARM instead of VIMCO at matched computation | Linear Bernoulli VAE · Omniglot, train multi-sample bound | Multi-sample bound: -117.68 ± 0.07 -> -116.70 ± 0.03 | Helped |
| Use DisARM instead of VIMCO | Bernoulli VAE · Multi-sample variational bounds across datasets and settings | Multi-sample bound | Mixed |
| Use DisARM instead of RELAX | Bernoulli VAE · Hierarchical models across datasets | Train ELBO | Mixed |
Closest papers by meaning
- ARMS: Antithetic-REINFORCE-Multi-Sample Gradient for Binary Variables (ICML 2021 · similarity 0.92)
- Learning Latent Variable Models via Jarzynski-adjusted Langevin Algorithm (NeurIPS 2025 · similarity 0.84)
- TUBE: Tangent Upper Bound on Evidence for Discrete Diffusion Language Models (NeurIPS 2026 · similarity 0.83)
- Outsourced Diffusion Sampling: Efficient Posterior Inference in Latent Spaces of Generative Models (ICML 2025 · similarity 0.83)
- Distribution-Conditioned Adversarial Variational Autoencoder for Valid Instrumental Variable Generation (AAAI 2024 · similarity 0.83)
- DisCo-Diff: Enhancing Continuous Diffusion Models with Discrete Latents (ICML 2024 · similarity 0.83)
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.