Selective Jacobi Decoding Speeds Up Discrete Normalizing Flows
A new method, Selective Jacobi Decoding, accelerates inference in discrete autoregressive normalizing flows by relaxing strict sequential dependencies. The approach identifies that high-quality samples can be generated without conditioning on all preceding sub-variables, and that low dependency redundancy occurs in initial layers. This enables parallel computation, significantly reducing generation time while maintaining quality. The paper is available on arXiv (2505.24791).
Key facts
- Selective Jacobi Decoding accelerates inference in discrete autoregressive normalizing flows.
- Strict sequential dependency in inference is unnecessary for high-quality samples.
- Sub-variables can be approximated without conditioning on all preceding sub-variables.
- Models exhibit low dependency redundancy in initial layers and higher redundancy later.
- The method enables parallel computation during inference.
- The paper is available on arXiv with ID 2505.24791.
- Discrete normalizing flows offer analytical log-likelihood computation and end-to-end training.
- Autoregressive modeling enhances expressive power but restricts parallel computation.
Entities
Institutions
- arXiv