Research map / Probabilistic and scientific modeling / Variational and generative sampling
ARSM: Augment-REINFORCE-Swap-Merge Estimator for Gradient Backpropagation Through Categorical 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 |
|---|---|---|---|
| Use ARSM instead of REINFORCE | Single categorical hidden-layer VAE · Binarized MNIST, training | negative ELBO (nats): 127.0 vs 82.0 | Helped |
| Use ARSM instead of ST Gumbel-Softmax | Single categorical hidden-layer VAE · Binarized MNIST, testing | negative ELBO (nats): 96.4 vs 86.7 | Helped |
| Use ARSM instead of Gumbel-Softmax | Stochastic categorical network · MNIST conditional density estimation, test | negative log-likelihood: 59.7 vs 58.3 ± 0.2 | Helped |
| Add the merge step to ARS (ablation) | Categorical VAE · Binarized MNIST, training | negative ELBO (nats): 97.4 -> 82.0 | Helped |
| Add swapping and merging to AR (ablation) | Categorical VAE · Binarized MNIST, training | negative ELBO (nats): 133.6 -> 82.0 | Helped |
| Add a second categorical hidden layer (ablation) | ARSM categorical VAE · Binarized MNIST, training | negative ELBO (nats): 82.0 -> 78.3 | Helped |
| Add a second categorical hidden layer (ablation) | ARSM categorical VAE · Binarized MNIST, testing | negative ELBO (nats): 86.7 -> 89.5 | Hurt |
| Use ARSM policy gradients with pseudo-action rollouts instead of A2C or RELAX | Discrete-action policy optimization · Cart Pole, Acrobot, and Lunar Lander | moving-average reward | Helped |
Closest papers by meaning
- ARM: Augment-REINFORCE-Merge Gradient for Stochastic Binary Networks (ICLR 2019 · similarity 0.88)
- CARMS: Categorical-Antithetic-REINFORCE Multi-Sample Gradient Estimator (NeurIPS 2021 · similarity 0.87)
- Gradient Estimation for Binary Latent Variables via Gradient Variance Clipping (AAAI 2023 · similarity 0.87)
- Backpropagation through the Void: Optimizing control variates for black-box gradient estimation (ICLR 2018 · similarity 0.86)
- DisARM: An Antithetic Gradient Estimator for Binary Latent Variables (NeurIPS 2020 · similarity 0.86)
- Estimating Gradients for Discrete Random Variables by Sampling without Replacement (ICLR 2020 · 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.