Gaussianization Flows
This work addresses the need for efficient and invertible flow models in machine learning, offering improvements in expressivity and robustness, though it is incremental relative to existing flow methods.
The authors tackled the problem of transforming any continuous random vector into a Gaussian distribution using iterative Gaussianization, resulting in a new normalizing flow model that achieves better or comparable performance on tabular datasets compared to existing models like Real NVP, Glow, and FFJORD, with easier initialization and better robustness.
Iterative Gaussianization is a fixed-point iteration procedure that can transform any continuous random vector into a Gaussian one. Based on iterative Gaussianization, we propose a new type of normalizing flow model that enables both efficient computation of likelihoods and efficient inversion for sample generation. We demonstrate that these models, named Gaussianization flows, are universal approximators for continuous probability distributions under some regularity conditions. Because of this guaranteed expressivity, they can capture multimodal target distributions without compromising the efficiency of sample generation. Experimentally, we show that Gaussianization flows achieve better or comparable performance on several tabular datasets compared to other efficiently invertible flow models such as Real NVP, Glow and FFJORD. In particular, Gaussianization flows are easier to initialize, demonstrate better robustness with respect to different transformations of the training data, and generalize better on small training sets.