Research map / Probabilistic and scientific modeling / Variational and generative sampling
Bidirectional Helmholtz Machines
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 the top-down and bottom-up models as a geometric-mean bidirectional Helmholtz machine | BiHM · MUSHROOMS | NLL: 9.55 -> 9.40 | Helped |
| Train the top-down and bottom-up models as a geometric-mean bidirectional Helmholtz machine | BiHM · ADULT | NLL: 13.16 -> 13.82 | Hurt |
| Train the top-down and bottom-up models as a geometric-mean bidirectional Helmholtz machine | BiHM · CONNECT4 | NLL: 11.68 -> 12.31 | Hurt |
| Train the top-down and bottom-up models as a geometric-mean bidirectional Helmholtz machine | BiHM · DNA | NLL: 81.04 -> 86.92 | Hurt |
| Train the top-down and bottom-up models as a geometric-mean bidirectional Helmholtz machine | BiHM · NIPS-0-12 | NLL: 271.11 -> 272.71 | Hurt |
| Train the top-down and bottom-up models as a geometric-mean bidirectional Helmholtz machine | BiHM · OCR-LETTERS | NLL: 26.43 -> 27.30 | Hurt |
| Train the top-down and bottom-up models as a geometric-mean bidirectional Helmholtz machine | BiHM · RCV1 | NLL: 46.10 -> 46.98 | Hurt |
| Train the top-down and bottom-up models as a geometric-mean bidirectional Helmholtz machine | BiHM · WEB | NLL: 27.92 -> 28.22 | Hurt |
| Use a 12-latent-layer BiHM rather than RWS with 5 latent layers | MNIST generative model · binarized MNIST test | -log p(x) estimate: 85.4 -> 84.3 | Helped |
| Use a 12-latent-layer BiHM rather than IWAE with 2 latent layers | MNIST generative model · binarized MNIST test | -log p(x) estimate: 85.3 -> 84.3 | Helped |
| Use the trained BiHM proposal rather than the RWS proposal | importance sampling on MNIST · MNIST test | ESS: 0.10% ± 0.06 -> 11.9% ± 1.1 | Helped |
| Leave out individual latent layers from the full model (ablation) | BiHM · binarized MNIST | test log-likelihood | Hurt |
Closest papers by meaning
- Flexible and accurate inference and learning for deep generative models (NeurIPS 2018 · similarity 0.93)
- Ladder Variational Autoencoders (NeurIPS 2016 · similarity 0.86)
- Variational Laplace Autoencoders (ICML 2019 · similarity 0.86)
- From Variational to Deterministic Autoencoders (ICLR 2020 · similarity 0.86)
- Multilayer Matrix Factorization via Dimension-Reducing Diffusion Variational Inference (ICML 2025 · similarity 0.84)
- Unified Generation, Reconstruction, and Representation: Generalized Diffusion with Adaptive Latent Encoding-Decoding (ICML 2024 · similarity 0.84)
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.