Research map / Probabilistic and scientific modeling / Variational and generative sampling
Bit-Swap: Recursive Bits-Back Coding for Lossless Compression with Hierarchical 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 |
|---|---|---|---|
| Replace BB-ANS with recursive Bit-Swap coding | Hierarchical latent variable model · MNIST | compression rate (bits/dim): 1.48 -> 1.29 | Helped |
| Replace BB-ANS with recursive Bit-Swap coding | Hierarchical latent variable model · CIFAR-10 | compression rate (bits/dim): 4.19 -> 3.82 | Helped |
| Replace BB-ANS with recursive Bit-Swap coding | Hierarchical latent variable model · ImageNet (32×32) | compression rate (bits/dim): 4.66 -> 4.50 | Helped |
| Replace BB-ANS with recursive Bit-Swap coding | Hierarchical latent variable model · ImageNet (unscaled & cropped) | compression rate (bits/dim): 3.62 -> 3.51 | Helped |
| Use Bit-Swap rather than BB-ANS | Depth-8 hierarchical latent variable model · MNIST test | initial bitrate (bits/dim): 21.93 ± 0.34 -> 3.34 ± 0.33 | Helped |
| Use Bit-Swap rather than BB-ANS | Depth-8 hierarchical latent variable model · CIFAR-10 test | initial bitrate (bits/dim): 44.24 ± 0.87 -> 6.53 ± 0.74 | Helped |
| Increase latent depth from 1 to 8 (ablation) | Hierarchical latent variable model · MNIST | negative ELBO (bits/dim): 1.35 -> 1.27 | Helped |
| Increase latent depth from 1 to 8 (ablation) | Hierarchical latent variable model · CIFAR-10 | negative ELBO (bits/dim): 4.57 -> 3.78 | Helped |
| Increase latent depth from 1 to 4 (ablation) | Hierarchical latent variable model · ImageNet (32×32) | negative ELBO (bits/dim): 4.94 -> 4.48 | Helped |
| Increase latent depth from 2 to 8 (ablation) | BB-ANS · MNIST test | CMA compression rate at n=100 (bits/dim): 1.33 ± 0.03 -> 1.48 ± 0.03 | Hurt |
| Increase latent depth from 2 to 8 (ablation) | Bit-Swap · MNIST test | CMA compression rate at n=100 (bits/dim): 1.30 ± 0.03 -> 1.29 ± 0.03 | No clear effect |
| Increase latent depth from 2 to 8 (ablation) | BB-ANS · CIFAR-10 test | CMA compression rate at n=100 (bits/dim): 3.93 ± 0.08 -> 4.19 ± 0.09 | Hurt |
Closest papers by meaning
- Practical lossless compression with latent variables using bits back coding (ICLR 2019 · similarity 0.92)
- Improving Lossless Compression Rates via Monte Carlo Bits-Back Coding (ICML 2021 · similarity 0.88)
- Learned Lossless Image Compression based on Bit Plane Slicing (CVPR 2024 · similarity 0.84)
- Progressive Compression with Universally Quantized Diffusion Models (ICLR 2025 · similarity 0.83)
- BitStack: Any-Size Compression of Large Language Models in Variable Memory Environments (ICLR 2025 · similarity 0.83)
- LittleBit: Ultra Low-Bit Quantization via Latent Factorization (NeurIPS 2025 · 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.