Research map / Probabilistic and scientific modeling / Variational and generative sampling
Compression with Flows via Local Bits-Back Coding
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 local bits-back coding with a Flow++ model | Flow++ · CIFAR10, ImageNet 32x32, and ImageNet 64x64 test sets | bits/dim: 3.116 -> 3.118; 3.871 -> 3.875; 3.701 -> 3.703 | No clear effect |
| Use the compositional algorithm instead of the black-box algorithm | Flow++ · CIFAR10 | encoding time (seconds/datapoint): 64.37 ± 1.05 -> 0.77 ± 0.01 | Helped |
| Use the compositional algorithm instead of the black-box algorithm | Flow++ · ImageNet 32x32 | encoding time (seconds/datapoint): 534.74 ± 5.91 -> 0.93 ± 0.02 | Helped |
| Use the compositional algorithm instead of the black-box algorithm | Flow++ · ImageNet 64x64 | encoding time (seconds/datapoint): 1349.65 ± 2.30 -> 0.69 ± 0.02 | Helped |
| Decrease discretization volume while keeping noise relatively large | Local bits-back · CIFAR10 subset | net codelength (bits/dim): At σ = 2⁻⁶, δ^(1/d) = 2⁻³²: 4.520 ± 0.082; δ^(1/d) = 2⁻⁸: 17.040 ± 0.332 | No clear effect |
| Reduce noise and use sufficiently fine discretization | Local bits-back · CIFAR10 subset | net codelength (bits/dim): 4.520 ± 0.082 -> 3.099 ± 0.140 | Helped |
| Shrink discretization volume | Local bits-back · CIFAR10 subset | auxiliary bits required: At σ = 2⁻¹⁴, δ^(1/d) = 2⁻³²: 51.840 ± 0.088; δ^(1/d) = 2⁻⁸: 10.134 ± 0.000 | Hurt |
Closest papers by meaning
- Practical lossless compression with latent variables using bits back coding (ICLR 2019 · similarity 0.89)
- Integer Discrete Flows and Lossless Compression (NeurIPS 2019 · similarity 0.89)
- Joint Autoregressive and Hierarchical Priors for Learned Image Compression (NeurIPS 2018 · similarity 0.88)
- Progressive Compression with Universally Quantized Diffusion Models (ICLR 2025 · similarity 0.87)
- Analysis of Learning a Flow-based Generative Model from Limited Sample Complexity (ICLR 2024 · similarity 0.85)
- Compression of Structured Data with Autoencoders: Provable Benefit of Nonlinearities and Depth (ICML 2024 · similarity 0.85)
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.