Residual Flows for Invertible Generative Modeling

TL;DR

Residual Flows introduce an improved invertible residual network framework with unbiased density estimation, reduced memory usage, and enhanced activation functions, achieving state-of-the-art results in flow-based generative modeling.

Contribution

The paper presents Residual Flows, a novel invertible residual network approach with unbiased log-density estimation and improved activation functions, advancing flow-based generative modeling.

Findings

Achieves state-of-the-art density estimation performance.

Outperforms coupling block networks in generative and discriminative tasks.

Reduces training memory requirements through an alternative gradient series.

Abstract

Flow-based generative models parameterize probability distributions through an invertible transformation and can be trained by maximum likelihood. Invertible residual networks provide a flexible family of transformations where only Lipschitz conditions rather than strict architectural constraints are needed for enforcing invertibility. However, prior work trained invertible residual networks for density estimation by relying on biased log-density estimates whose bias increased with the network's expressiveness. We give a tractable unbiased estimate of the log density using a "Russian roulette" estimator, and reduce the memory required during training by using an alternative infinite series for the gradient. Furthermore, we improve invertible residual blocks by proposing the use of activation functions that avoid derivative saturation and generalizing the Lipschitz condition to induced mixed norms. The resulting approach, called Residual Flows, achieves state-of-the-art performance on density estimation amongst flow-based models, and outperforms networks that use coupling blocks at joint generative and discriminative modeling.