Discrete Flows: Invertible Generative Models of Discrete Data
This work addresses the challenge of applying generative models to discrete data, which is incremental by extending continuous flow methods to discrete domains.
The paper tackles the problem of modeling discrete distributions with normalizing flows, showing that flows can be extended to discrete events using a simple change-of-variables formula without log-determinant-Jacobian computations. It demonstrates that discrete autoregressive flows outperform baselines on synthetic tasks and Potts models, and bipartite flows achieve competitive performance on character-level language modeling datasets like Penn Tree Bank and text8.
While normalizing flows have led to significant advances in modeling high-dimensional continuous distributions, their applicability to discrete distributions remains unknown. In this paper, we show that flows can in fact be extended to discrete events---and under a simple change-of-variables formula not requiring log-determinant-Jacobian computations. Discrete flows have numerous applications. We consider two flow architectures: discrete autoregressive flows that enable bidirectionality, allowing, for example, tokens in text to depend on both left-to-right and right-to-left contexts in an exact language model; and discrete bipartite flows that enable efficient non-autoregressive generation as in RealNVP. Empirically, we find that discrete autoregressive flows outperform autoregressive baselines on synthetic discrete distributions, an addition task, and Potts models; and bipartite flows can obtain competitive performance with autoregressive baselines on character-level language modeling for Penn Tree Bank and text8.